Paper deep dive
Med-AR: Autoregressive Vision-Language Pretraining for Long-Tailed Chest X-Ray Classification and Uncertainty-Aware Evaluation
Janhavi Prabhu, Sahil, Akshay V, Shivam Shukla, Manoj Tadepalli, Preetham Putha
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 96%
Last extracted: 9/25/2026, 5:02:14 AM
Summary
This paper introduces Med-AR, a radiology-native autoregressive vision-language model (Med-AR-8B and Med-AR-2B) designed for long-tailed chest X-ray classification. The authors compare Med-AR against contrastive pretraining baselines like Med-CLIP, CheXFound, EVA-Base, ARK, and BioViL-T. Using a shared InternViT backbone and ML-Decoder head, they demonstrate that Med-AR-8B outperforms Med-CLIP in mean AUROC and AUPRC across head, medium, and tail findings on PadChest, MIMIC-CXR, and CheXpert. Med-AR also shows improved selective-prediction performance with lower excess area under the risk-coverage curve. The study highlights the benefits of autoregressive pretraining with heterogeneous supervision (structured reports, region annotations) for capturing fine-grained and rare abnormalities.
Entities (16)
Relation Signals (10)
Med-AR-8B → evaluatedon → PadChest
confidence 97% · Across PadChest, MIMIC-CXR, and CheXpert
Med-AR-8B → evaluatedon → CheXpert
confidence 97% · Across PadChest, MIMIC-CXR, and CheXpert
Med-AR-8B → evaluatedon → MIMIC-CXR
confidence 97% · Across PadChest, MIMIC-CXR, and CheXpert
Med-AR-8B → usesbackbone → InternViT
confidence 96% · using the InternViT architecture trained under each objective
Med-CLIP → usesbackbone → InternViT
confidence 96% · Our contrastive baseline combines InternViT with BiomedVLP-CXR-BERT
Med-AR-8B → outperforms → Med-CLIP
confidence 95% · Med-AR-8B outperforms Med-CLIP in mean AUROC and AUPRC for head, medium, and tail findings.
Med-AR-8B → usesclassificationhead → ML-Decoder
confidence 95% · using a common ML-Decoder classification head
Med-AR-8B → improvesmetricon → AUPRC
confidence 94% · On MIMIC-CXR, it increases tail-label mean AUPRC from 0.1033 to 0.1441.
Med-AR → achieveslowermetricthan →
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Long-tailed chest X-ray classification requires visual representations that capture both common abnormalities and subtle, infrequent findings. We propose Med-AR-8B and Med-AR-2B, two radiology-native autoregressive vision-language models pretrained with structured reports, abnormality-focused text, and region annotations. We evaluate the transfer of their visual encoders to multi-label classification against contrastive, self-supervised, and supervised pretrained encoders, including Med-CLIP, CheXFound, EVA-Base, ARK, and BioViL-T, using a common ML-Decoder classification head. To assess fine-grained recognition, we also construct LLM-expanded, report-derived label sets for MIMIC-CXR and CheXpert. Across PadChest, MIMIC-CXR, and CheXpert, Med-AR-8B outperforms Med-CLIP in mean AUROC and AUPRC for head, medium, and tail findings. On MIMIC-CXR, it increases tail-label mean AUPRC from 0.1033 to 0.1441. Med-AR-2B achieves the strongest discrimination results on PadChest. Across the broader encoder comparison, a Med-AR variant achieves the highest mean AUROC and AUPRC in every reported prevalence group on each public dataset. Both Med-AR variants also achieve lower excess area under the risk-coverage curve than Med-CLIP on all three public datasets, indicating improved selective-prediction performance under the evaluated protocol. Internal results are metric-dependent, with Med-CLIP retaining advantages in overall and tail AUPRC and in selective prediction. These findings establish Med-AR as a strong pretraining recipe for long-tailed chest X-ray classification on the evaluated public benchmarks and demonstrate the value of assessing discrimination and selective prediction together.
Tags
Links
- Source: https://arxiv.org/abs/2609.29156v1
- Canonical: https://arxiv.org/abs/2609.29156v1
Trouble viewing inline? Open PDF directly →
Full Text
233,910 characters extracted from source content.
Expand or collapse full text
Med-AR: Autoregressive Vision-Language Pretraining for Long-Tailed Chest X-Ray Classification and Uncertainty-Aware Evaluation Janhavi Prabhu, Sahil, Akshay V, Shivam Shukla, Manoj Tadepalli, and Preetham Putha Qure.ai, Bangalore, India janhavi.prabhu@qure.ai sahil@qure.ai Akshay.V@qure.ai shivam.shukla@qure.ai manoj.tadepalli@qure.ai preetham.putha@qure.ai Abstract. Long-tailed chest X-ray classification requires visual rep- resentations that capture both common abnormalities and subtle, infrequent findings. We propose Med-AR-8B and Med-AR-2B, two radiology-native autoregressive vision-language models pretrained with structured reports, abnormality-focused text, and region annotations. We evaluate the transfer of their visual encoders to multi-label classi- fication against contrastive, self-supervised, and supervised pretrained encoders, including Med-CLIP, CheXFound, EVA-Base, ARK, and BioViL-T, using a common ML-Decoder classification head. To assess fine-grained recognition, we also construct LLM-expanded, report- derived label sets for MIMIC-CXR and CheXpert. Across PadChest, MIMIC-CXR, and CheXpert, Med-AR-8B outperforms Med-CLIP in mean AUROC and AUPRC for head, medium, and tail findings. On MIMIC-CXR, it increases tail-label mean AUPRC from 0.1033 to 0.1441. Med-AR-2B achieves the strongest discrimination results on PadChest. Across the broader encoder comparison, a Med-AR variant achieves the highest mean AUROC and AUPRC in every reported prevalence group on each public dataset. Both Med-AR variants also achieve lower excess area under the risk–coverage curve than Med-CLIP on all three public datasets, indicating improved selective-prediction performance under the evaluated protocol. Internal results are metric-dependent, with Med- CLIP retaining advantages in overall and tail AUPRC and in selective prediction. These findings establish Med-AR as a strong pretraining recipe for long-tailed chest X-ray classification on the evaluated public benchmarks and demonstrate the value of assessing discrimination and selective prediction together. Keywords: MedCLIP· Vision-Language Models· Long-Tailed Classi- fication· Uncertainty· Autoregressive Pretraining· Chest X-ray arXiv:2609.29156v1 [cs.CV] 24 Sep 2026 2Prabhu et al. 1 Introduction Chest X-rays (CXRs) are among the most frequently performed imaging studies for screening, triage, and longitudinal monitoring of thoracic disease, making robust automated abnormality recognition clinically important. However, multi- label CXR classification remains challenging because thoracic findings exhibit a highly long-tailed distribution: a small subset of common abnormalities accounts for the majority of positive samples, while many clinically significant findings appear only infrequently [16,22]. Consequently, models may achieve strong ag- gregate metrics yet fail on rare conditions, where reliable performance is often most clinically necessary. This problem is further complicated by disease co- occurrence, reporting variability, label noise, and dataset heterogeneity. Rare abnormality recognition in CXRs also differs from standard image-level classification. Many uncommon or clinically subtle findings are spatially local- ized, low contrast, or interpretable only in relation to surrounding anatomical structures. Although broad abnormalities such as cardiomegaly or large pleu- ral effusions can often be identified from coarse global structure, other findings require fine-grained local evidence together with anatomical and contextual un- derstanding. Effective long-tail CXR classification therefore depends on repre- sentations that preserve localized visual detail while simultaneously maintaining clinically meaningful global context, rather than relying solely on pooled image summaries [17,47]. To address these challenges, recent work has increasingly explored vision- language pretraining using paired radiology reports as supervision. Contrastive approaches, including CLIP, CXR-CLIP, Language over Labels, and MedCLIP, learn transferable visual representations by aligning image and text embed- dings [33,51,45,41]. These methods have demonstrated strong performance in medical imaging [4,17,47]. Global image–text alignment compresses report infor- mation into pooled representations, potentially limiting the supervision available for fine-grained findings. This limitation does not apply uniformly to contrastive learning: local and knowledge-guided approaches explicitly model finer-grained relationships. Our comparison focuses on the contrastive recipe evaluated here, rather than all possible contrastive formulations. Autoregressive vision-language pretraining provides an alternative supervi- sion paradigm. Rather than optimizing only global image-text agreement, au- toregressive models condition on the image and generate textual outputs sequen- tially token by token [30,6,2,7,55,25]. This formulation naturally supports richer and more heterogeneous supervision from the same study, including full reports, impression summaries, structured findings, region descriptions, bounding-box annotations, question answering, and step-by-step clinical reasoning. Because su- pervision is provided through token-level prediction rather than a single pooled similarity objective, rare findings contribute explicit gradients whenever they ap- pear, potentially encouraging representations that better preserve fine-grained anatomical and contextual information. Autoregressive training therefore offers a scalable framework for integrating diverse forms of clinical supervision within a unified objective, without requiring separate task-specific formulations. In ra- Med-AR for Long-Tailed Chest X-Ray Classification3 diology, where interpretation depends not only on identifying abnormalities but also on understanding their anatomical location, prominence, and contextual relationships, multimodal next-token prediction may therefore produce more transferable visual representations for downstream long-tail recognition. Despite growing interest in large autoregressive vision-language models, it re- mains unclear whether radiology-native autoregressive pretraining provides bet- ter downstream transfer than medical contrastive pretraining under controlled experimental conditions. Prior comparisons often differ simultaneously in archi- tecture, pretraining corpus, optimization strategy, and downstream evaluation protocol, making it difficult to isolate the contribution of the pretraining ob- jective itself. Furthermore, evaluations of medical vision-language models have primarily focused on aggregate discrimination performance, with comparatively less attention given to long-tail behavior and uncertainty-aware evaluation, both of which are critical for reliable clinical deployment [11,56,18,14]. In this work, we address this gap through a comparison of contrastive and autoregressive pretraining under a matched downstream architecture for long- tailed multi-label CXR classification. We compare MedCLIP-style contrastive pretraining and autoregressive vision-language pretraining using the InternViT architecture trained under each objective, followed by independent downstream fine-tuning with an ML-Decoder classification head, thereby controlling the downstream architecture while comparing complete pretraining recipes. Our autoregressive framework is trained in a radiology-native setting using heteroge- neous supervision, including structured multi-section report templates, auxiliary abnormality-focused targets, and approximately 100k region annotations. To evaluate rare-finding transfer more rigorously, we additionally construct LLM- expanded label sets for MIMIC-CXR and CheXpert that increase the granularity of downstream evaluation. Experiments across an internal dataset, MIMIC- CXR, CheXpert, and PadChest demonstrate that autoregressive pretraining achieves stronger downstream discrimination on public benchmarks, improves performance on multiple tail abnormalities, and yields lower EAURC on the public datasets, while the contrastive baseline retains lower EAURC on the internal dataset [44,14]. Our contributions are threefold: – We compare contrastive and autoregressive pretraining recipes for long-tailed chest X-ray classification, using separate InternViT encoders independently fine-tuned with the same ML-Decoder architecture. – We develop a radiology-native autoregressive pretraining framework that in- tegrates heterogeneous supervision from structured reports, auxiliary text targets, and region-level annotations, enabling the encoder to learn repre- sentations beyond global image-report alignment. – We provide a comprehensive evaluation across internal and public CXR datasets, including head-tail analysis, LLM-expanded long-tail benchmarks for MIMIC-CXR and CheXpert, and uncertainty-aware assessment using risk-coverage metrics. 4Prabhu et al. 2 Related Work 2.1 Long-tailed chest X-ray classification Long-tailed recognition in chest X-rays has received increasing attention, par- ticularly following the introduction of the CXR-LT benchmark and challenge, which demonstrated that strong aggregate performance can conceal poor recog- nition of rare findings [15,16]. Existing approaches primarily address imbalance during supervised downstream training through techniques such as re-weighted losses, asymmetric objectives, resampling, data augmentation, and architectural modifications including multi-view fusion [54,36,31,24]. Although these methods improve classification under skewed label distributions, they focus largely on op- timization at the downstream stage rather than on the representations learned during pretraining. This limitation is particularly important in chest radiography, where rare ab- normalities are often subtle, spatially localized, and difficult to distinguish using coarse global features alone. Consequently, there has been growing interest in leveraging radiology reports as an additional source of supervision, since textual descriptions provide clinically meaningful context that may help models learn more transferable representations for rare and fine-grained findings. This moti- vation has led to the increasing adoption of vision-language pretraining methods in medical imaging. 2.2 Contrastive pretraining Medical vision-language pretraining commonly uses radiology reports as weak supervision for learning transferable visual representations. Early and widely adopted methods follow CLIP-style contrastive alignment, including CXR-CLIP, Language over Labels, and MedCLIP, which learn a shared image–text embed- ding space from paired or weakly paired image-report studies [33,51,45,41]. By incorporating report semantics during pretraining, these approaches improve representation quality beyond purely supervised image classification. Building on this paradigm, subsequent work has explored richer forms of alignment to better capture the structure and granularity of radiology reports. For example, GLoRIA introduces both global and local alignment objectives be- tween report words and image regions, while MedKLIP incorporates knowledge- enhanced triplet encoding derived from medical reports to guide visual learning from chest X-rays [17,4,47]. Similar trends have also emerged outside medical imaging, where methods such as DenseCLIP and FineCLIP improve dense and fine-grained recognition by modeling local visual–textual correspondences rather than relying solely on pooled embeddings [34,20]. In parallel, recent studies use large language models to extract structured disease information from reports, improving weak supervision and report understanding [43]. Collectively, these works suggest that richer supervision and finer-grained image-text interactions can improve representation learning. Global alignment remains a common component, although local and knowledge-guided contrastive Med-AR for Long-Tailed Chest X-Ray Classification5 methods also exploit finer-grained supervision. This has motivated growing in- terest in autoregressive vision-language modeling, where supervision is provided through sequential next-token prediction rather than pooled embedding align- ment alone. 2.3 Autoregressive pretraining Autoregressive and generative objectives have emerged as a strong alternative for multimodal representation learning. Early work such as VirTex demonstrated that generating captions from semantically rich annotations can produce trans- ferable visual features, while SimVLM introduced a simplified large-scale pre- training framework based on prefix language modeling [9,42]. Subsequent meth- ods further expanded this paradigm. CoCa combines contrastive alignment with autoregressive caption generation within a unified framework, and Emu ex- tends next-token prediction to interleaved image–text sequences [52,40]. More recently, AIM and AIMv2 showed that large autoregressively pretrained vision encoders can scale effectively and transfer well across classification, localization, and grounding tasks [10,12]. A key advantage of autoregressive supervision is that it naturally supports dense and heterogeneous training signals. Instead of optimizing a single global similarity objective, autoregressive models can learn from full reports, impres- sion summaries, structured findings, region descriptions, and other clinically rel- evant textual outputs. This property is particularly appealing in radiology, where understanding subtle abnormalities often depends on contextual and localized information. In the medical domain, generative pretraining has therefore been explored for joint image-text understanding and report generation [30], while recent multimodal systems such as Florence-VL and InternVL3 further demon- strate the increasing role of autoregressive vision-language modeling [6,55]. Despite this progress, there remains limited evidence on whether autoregres- sive pretraining provides advantages over medical contrastive pretraining for long-tailed multi-label chest X-ray classification under controlled experimental settings. In particular, prior comparisons often differ simultaneously in archi- tecture, optimization strategy, and downstream evaluation protocol, making it difficult to isolate the effect of the pretraining objective itself. 2.4 Uncertainty-aware evaluation in medical imaging Beyond discrimination performance alone, reliable deployment of medical AI sys- tems also requires robust uncertainty estimation. This is especially important in long-tailed classification settings, where rare findings are underrepresented dur- ing training and therefore more susceptible to unreliable predictions. Prior work in machine learning and radiology distinguishes between aleatoric uncertainty, which arises from inherent noise or ambiguity in the data, and epistemic un- certainty, which reflects uncertainty in the model parameters due to limited or biased observations [18,11]. In chest radiography, both forms of uncertainty are 6Prabhu et al. clinically relevant because findings may be visually ambiguous, sparsely repre- sented, or confounded by overlapping pathology. Recent work has increasingly emphasized that uncertainty should be eval- uated jointly with prediction reliability and clinical decision risk rather than treated as a secondary metric [56,38]. Beyond calibration alone, selective pre- diction and risk-coverage analysis assess whether a model’s confidence can ap- propriately rank its own errors, enabling uncertain predictions to be deferred to clinicians when necessary [14,13]. However, uncertainty-aware evaluation re- mains relatively underexplored in studies of medical vision-language pretraining, particularly in the long-tailed multi-label setting where rare diseases are both more difficult to recognize and more prone to overconfident failure modes. 3 Methodology This study investigates how different vision-language pretraining paradigms in- fluence representation learning and downstream transfer in long-tailed chest X- ray classification. The problem is particularly challenging because rare thoracic abnormalities are often subtle, sparsely represented, and frequently co-occur with more common findings [16,22]. In such settings, robust recognition depends on representations that preserve both localized visual detail and broader clinical context. The contrastive recipe considered here aligns pooled image and text representations, whereas autoregressive pretraining learns by sequentially pre- dicting clinically grounded text conditioned on the image. This autoregressive formulation naturally supports heterogeneous supervision signals, including re- port generation, structured findings, region descriptions, question answering, and bounding-box annotations within a unified training framework. If such supervi- sion improves representation quality, autoregressive pretraining may provide a scalable approach for incorporating increasingly diverse forms of clinical super- vision. To evaluate these differences, we perform a comparison under matched downstream architecture between contrastive and autoregressive pretraining un- der matched downstream evaluation settings, described in the following sections. 3.1 Overview We study the effect of the pretraining paradigm on long-tailed multi-label chest X-ray (CXR) classification by comparing two vision-language pretraining strate- gies under a controlled downstream setup: (i) MedCLIP-style contrastive pre- training [41], and (i) autoregressive pretraining using an InternVL3-style vision- language model [55]. To control the downstream architecture, both approaches use independently pretrained instances of the same image-backbone family and are fine-tuned with the same classification head, namely ML-Decoder [35] shown in Figure 1. Formally, let x denote a chest X-ray image and let y ∈ 0, 1 C denote the multi-label target vector over C abnormalities. In both settings, a pretrained Med-AR for Long-Tailed Chest X-Ray Classification7 Fig. 1: Overview of the proposed framework. Top: In autoregressive pretraining, each chest X-ray is dynamically tiled at high resolution according to a matched aspect ratio and encoded by InternViT-300M. The visual features are trans- formed into vision tokens through pixel unshuffle and an MLP projector, then combined with prompt-conditioned text tokens in a language decoder. The model is trained with heterogeneous radiology-native supervision, including report gen- eration, medical reasoning, structured abnormality outputs, and region-level de- scriptions such as abnormality bounding boxes. Bottom: For downstream long- tailed multi-label classification, the pretrained InternViT encoder is transferred and coupled with an ML-Decoder head, which predicts logits for all abnormali- ties and is optimized with weighted asymmetric loss. 8Prabhu et al. vision encoder E(·) first maps the image to a set of visual features, z = E(x),(1) which are then passed to independently trained instances of the same multi-label classifier architecture g(·) to obtain per-label probabilities, ˆ p = g(z)∈ [0, 1] C .(2) The transferred encoders share an architecture but are pretrained with different objectives and supervision. In particular, the autoregressive recipe additionally uses structured and region-level targets. This design controls the downstream architecture but does not isolate the effect of the pretraining objective from the supervision, initialization, or optimization recipe. 3.2 Shared Vision Backbone Architecture For both paradigms, we use InternViT-300M as the visual backbone. Intern- ViT is a transformer-based image encoder designed for high-resolution visual understanding and defines the common encoder architecture in our comparison. In the autoregressive setup, the encoder operates within the InternVL3 frame- work [55]; in the contrastive setup, an independent encoder is trained using our MedCLIP-style contrastive recipe [41]. By transferring the same encoder family to the downstream task, we reduce architectural confounding while comparing the transfer performance of the two recipes. 3.3 Autoregressive Vision-Language Pretraining Our autoregressive pretraining setup is implemented using InternVL3 [55], as illustrated at the top of Fig. 1, where an InternViT image encoder is trained jointly with a language decoder via conditional next-token prediction. Given an image x and a textual prompt q, the model generates a target sequence t = (t 1 ,t 2 ,...,t T ),(3) which may correspond to a structured report, an impression-style summary, a compact abnormality list, or a region-grounded textual description. The training objective maximizes the conditional likelihood of the target sequence: L AR =− T X k=1 logp(t k | x,q,t <k ).(4) This objective couples visual representation learning with language genera- tion. The image is encoded into a sequence of visual tokens that condition the decoder, and the encoder is trained to produce features that support accurate and coherent text generation. This motivates the hypothesis that the learned Med-AR for Long-Tailed Chest X-Ray Classification9 representations retain structured and fine-grained information; classification re- sults alone do not directly establish visual grounding. A key advantage of this formulation is that it naturally supports multiple forms of supervision through prompt-conditioned generation. The same image can be used to generate different clinically meaningful outputs, such as detailed findings, concise impressions, abnormality summaries, or region-level descrip- tions. This provides dense, token-level supervision across multiple views of the data, rather than relying on a single global alignment signal. To further enrich supervision, we construct diverse training targets from ra- diology reports and auxiliary annotations. These include: (i) section-wise re- port templates, (i) short impression-like summaries, (i) compact abnormality- focused outputs, and (iv) region-level descriptions derived from bounding-box annotations. Such heterogeneous supervision is intended to associate localized image patterns with corresponding text, potentially benefiting subtle and spa- tially specific abnormalities. To study how the strength of the language modeling objective affects the learned visual features, we pretrain two models with identical InternViT en- coders but different decoder scales. One model uses a 2B-parameter language decoder (Med-AR-2B), while the other uses an 8B-parameter decoder (Med- AR-8B). A larger decoder can model more complex linguistic patterns, but may also dominate training or overfit to dataset-specific text, potentially limiting the generality of the visual features. Whether decoder capacity changes the infor- mation retained by the vision encoder remains an empirical question. After pretraining, the language decoder is discarded and only the vision en- coder is transferred to the downstream classification task. Both Med-AR-2B and Med-AR-8B are evaluated under an identical fine-tuning setup and compared against an InternViT encoder pretrained with MedCLIP, enabling a comparison of the resulting encoders under matched downstream architecture. 3.4 MedCLIP-Style Contrastive Pretraining Our contrastive baseline combines InternViT with BiomedVLP-CXR-BERT and uses the MedCLIP semantic-matching objective [41,3]. We retain the name Med-CLIP in the result tables for this implementation. Let v = f img (x) and u = f text (r) denote image and report embeddings. Medical semantic informa- tion from report-derived labels determines soft matching targets for image–text pairs, rather than assigning a positive target only to the paired report and treat- ing every other report as a negative. The bidirectional semantic-matching objective is L MedCLIP = 1 2 (L img→text +L text→img ),(5) where each directional term is a cross-entropy between the semantic matching targets and the similarity-based predicted matching distribution. This is distinct from standard paired InfoNCE: images and reports with related medical labels 10Prabhu et al. can receive nonzero matching targets even when they are not an original pair. After pretraining, the visual encoder is transferred to downstream classification. The baseline aligns pooled visual and textual representations with semantic supervision. The autoregressive recipe additionally uses structured generation targets and region annotations. The comparison therefore evaluates complete pretraining recipes rather than isolating the objective or establishing a limitation of all contrastive methods. 3.5 Downstream Fine-Tuning with ML-Decoder After pretraining, all three vision encoders are fine-tuned on downstream multi- label classification using ML-Decoder [35] as shown in bottom of Fig 1. ML- Decoder is a transformer-based prediction head designed for efficient large-scale multi-label recognition. Instead of learning an independent classifier for each label, it uses a fixed set of learnable queries that interact with the encoded image features to produce label predictions. Letz∈R N×d denote the output token sequence of the vision encoder, where N is the number of visual tokens and d is the feature dimension. ML-Decoder mapsz to class logits s = h(z)∈R C ,(6) followed by sigmoid activation to obtain class probabilities ˆp i = σ(s i ), i = 1,...,C.(7) A key advantage of ML-Decoder is that it scales efficiently to large label spaces while still modeling label dependencies and co-occurrence structure. This makes it particularly suitable for long-tailed CXR classification, where abnor- malities often co-occur and the number of classes can be large in fine-grained evaluation settings. 3.6 Loss Function for Long-Tailed Multi-Label Classification To address both inter-class and intra-class imbalance during downstream train- ing, we combine class-specific weighting with asymmetric loss (ASL) [36,24]. Class-specific weighting. Let y i ∈ 0, 1 denote the ground-truth label for class i, and let ρ i denote the positive sample ratio of class i in the training set. We define a class-aware sample weight as w i = y i exp(1− ρ i ) + (1− y i ) exp(ρ i ).(8) This weighting increases the contribution of rare positive classes while preserving the contribution of negative examples. Med-AR for Long-Tailed Chest X-Ray Classification11 Asymmetric loss. For each class i, the asymmetric loss is defined as L (i) ASL =−y i (1− ˆp i ) γ + log( ˆp i )− (1− y i ) ˆp γ − m,i log(1− ˆp m,i ),(9) where ˆp m,i = max( ˆp i − m, 0).(10) Here, γ + and γ − are focusing parameters for positive and negative samples, respectively, and m is a probability margin used to suppress easy negatives. Following prior long-tailed CXR classification work [24], we use γ + = 1, γ − = 4, and m = 0.05. Final training objective. The final downstream loss is computed as the weighted sum over classes: L cls = C X i=1 w i L (i) ASL (11) This objective jointly addresses inter-class imbalance, by up-weighting rare classes, and intra-class imbalance, by down-weighting easy negatives. As a result, it is well suited for long-tailed multi-label CXR classification. 3.7 Long-tailed evaluation protocol We study the downstream task as multi-label abnormality classification on four datasets: an internal clinical dataset, MIMIC-CXR, CheXpert and PadChest. This multi-dataset setup is important because long-tailed behavior can be sensi- tive to dataset composition, reporting style, disease prevalence, and annotation granularity. Evaluating on both internal and public datasets therefore helps dis- tinguish in-domain gains from broader transfer behavior. To make the long-tail analysis explicit, we partition labels separately within each dataset according to their frequency in the training split. Let c i denote the number of positive training examples for label i. Because the label-frequency distribution is strongly long-tailed, we first transform counts into log space: x i = log(1 + c i ).(12) We then compute dataset-specific thresholds from the empirical 33rd and 67th percentiles of the log-count distribution: T 1 = Q 0.33 (x), T 2 = Q 0.67 (x).(13) To obtain interpretable thresholds in the original count space, we map these quantiles back using the inverse transform: C 1 = exp(T 1 )− 1, C 2 = exp(T 2 )− 1.(14) Finding labels are then assigned to three groups directly according to their pos- itive training counts: tail : c i < C 1 ,medium : C 1 ≤ c i < C 2 ,head : c i ≥ C 2 .(15) 12Prabhu et al. This procedure provides a dataset-adaptive partition of the label space while applying the same rule across all benchmarks. Because the log transform preserves ordering, the groups principally reflect relative label-frequency ranks within each dataset, with boundary details determined by quantile interpo- lation and ties. Tail membership therefore does not imply the same absolute prevalence across datasets. The resulting head/medium/tail split enables finer- grained analysis of whether improvements from pretraining are concentrated on frequent labels or extend to intermediate- and low-frequency abnormalities. The dataset-specific count thresholds used in this work are summarized in Table 2. For MIMIC-CXR and CheXpert, we further expand the evaluation space through an LLM-based label extraction pipeline, producing a finer-grained long- tailed benchmark than the standard label set. This expanded label space, shown in Figure 2 and Figure 3, is not introduced as a separate task; rather, it serves as a more sensitive probe of representation quality under rare-label transfer, where differences between pretraining paradigms may be difficult to detect using only a small number of coarse disease categories. 4 Experiments 4.1 Dataset Preparation Pre-training: For pre-training we use an internal training dataset of 1.4 mil- lion chest X-rays and their corresponding reports, obtained from various hos- pitals, anonymized and cleaned. For MedCLIP pretraining, we leverage paired image–text data and use prompt engineering to design an effective instruction for extracting detailed disease labels from radiology reports, enabling fine-grained annotation of more than 150 findings, including rare abnormalities. The LLM (Qwen3 [49]) was prompted to produce a JSON object listing all present find- ings, guided by a curated list of over 150 finding categories. Downstream label expansion is described separately in Section 4.2. For autoregressive pre-training, we create a structured JSON template consisting of 15 subheadings covering the various anatomical regions to look at in a chest radiograph. Under each subhead- ing, we include further categorizations for abnormalities. In addition to the 15 headings, we also add Findings, Impression, Summary, Reasoning and Promi- nence Score sections. The prominence score section provides a dictionary of the report-derived findings along with a generated prominence score, from 0–5. We use Qwen3 to extract this information from the reports and output the response in the structured template format. In cases when an abnormality is not men- tioned or is not present, we instruct the LLM to use a negation statement. This procedure conflates non-mention with explicit absence and is a source of weak- supervision noise. The reasoning and prominence fields are generated from re- ports; they are not independently adjudicated visual assessments. Furthermore, we incorporate expert-annotated bounding boxes for approximately 20 thoracic findings (e.g., opacity, nodule, pneumothorax, mass), represented as four-point coordinates. For each supervision type: global report parsing, structured abnor- mality extraction, prominence scoring, and region-level bounding-box, we design Med-AR for Long-Tailed Chest X-Ray Classification13 dedicated prompt templates for Qwen3, ensuring consistent JSON outputs across these heterogeneous data sources which are then used for pretraining InternVL3 model. Finetuning: We conduct separate training experiments on four large-scale CXR datasets: an internal collection of chest radiographs (Internal), PadChest [5], CheXpert [19] and MIMIC-CXR [21]; the splits in Table 1 contain 1,872,398 training images and 297,608 test images in total. For each dataset, models are trained exclusively on its corresponding training split and evaluated on its own held-out test set, as summarized in Table 1. – Internal Dataset: The training split includes approximately 1.4M chest X- rays from multiple hospitals. An LLM fine-tuned on radiologist annotations extracts 40 report-derived labels, including hilar mass. – PadChest: Includes 160K CXRs labeled with 174 UMLS-derived findings. We use the 29 most frequent labels for training and the held-out split re- ported in Table 1. Volume loss is the least prevalent retained finding. Tail analysis is restricted to these retained labels, rather than the full PadChest vocabulary. – MIMIC-CXR: A public dataset with 173K CXRs (frontal) and reports. Our LLM extracts∼150 tags for granularity. We use 110 finding labels (with training prevalence > 0.0002) for training. Bullae is the least common re- tained finding. We plan to release the derived labels and split identifiers subject to the source dataset terms. – CheXpert: A public dataset with approximately 160K retained CXRs (frontal) and reports. Our LLM extracts ∼150 tags for granularity. We use 115 finding labels (with training prevalence > 0.0002) for training. Left ventricular assist device (LVAD) is the least common retained label. We plan to release the derived labels and split identifiers subject to the source dataset terms. DatasetTraining SamplesTesting Samples Internal1,432,373242,965 MIMIC-CXR152,24021,466 PadChest138,23522,626 CheXpert149,55010,551 Table 1: Dataset statistics, showing the number of training and testing samples for each dataset used in fine-tuning. 14Prabhu et al. Fig. 2: Distribution of extracted MIMIC-CXR tags. The classification label space contains 110 labels; available per-label results are listed in Supplementary Sec- tion S5. The data is split into head, medium, and tail tags. Train prevalence ≥ 0.02%; log-count thresholds: tail < 1106, medium < 4210, head ≥ 4210. Med-AR for Long-Tailed Chest X-Ray Classification15 Fig. 3: Distribution of the CheXpert tag extraction dataset with 115 tags. The data is split into head, medium, and tail tags. Train prevalence ≥ 0.02%; log- count thresholds: tail < 841, medium < 3519, head ≥ 3519. 16Prabhu et al. (a) Distribution of the Internal dataset with 40 tags. The data is split into head, medium, and tail tags. Train prevalence≥ 0.02%; log-count thresholds: tail < 3667, medium < 11176, head≥ 11176. (b) Distribution of the PadChest dataset with 29 tags. The data is split into head, medium, and tail tags. Train prevalence≥ 0.02%; log-count thresholds: tail < 2765, medium < 5361, head≥ 5361. Fig. 4: Tag distribution comparison for the Internal and PadChest datasets. Med-AR for Long-Tailed Chest X-Ray Classification17 The distribution of each tag in the train and test set is given in Figures 2 3 4a and 4b and training set thresholds for grouping into head, medium and tail for each dataset is mentioned in Table 2 and discussed in Sec 3.7. Table 2: Dataset-specific thresholds used to partition labels into head, medium, and tail groups based on the number of positive training examples. Thresholds are derived from the long-tailed label distribution of each dataset after filtering labels with training prevalence ≥ 0.02% as discussed in Section 3.7. DatasetTailMediumHead Internalc i < 3667 3667≤ c i < 11176 c i ≥ 11176 MIMIC-CXR c i < 1106 1106≤ c i < 4210c i ≥ 4210 PadChestc i < 2765 2765≤ c i < 5361c i ≥ 5361 CheXpertc i < 841841≤ c i < 3519c i ≥ 3519 4.2 LLM-based Label Extraction and Concordance Analysis To expand the label space of MIMIC-CXR and CheXpert, we derived structured labels from free-text radiology reports using a large language model (LLM). The prompt specified a comprehensive set of thoracic abnormalities and required the model to return their presence or absence in a structured JSON format. Positive findings were restricted to those supported by the report, with explicit instruc- tions to avoid inference beyond the documented text. We additionally included an other_findings field to capture abnormalities outside the predefined vo- cabulary. While this improved coverage, it introduced lexical variability across models, including alternative terminology and singular/plural forms; these free- form outputs were therefore mapped to a canonical label vocabulary during post-processing. We evaluated inter-model extraction concordance on a 1,000-report subset using seven LLMs: DeepSeek V4 Pro, Gemini 3.1 Pro, Gemini 3 Flash, GLM 5.1, Qwen3, Claude Haiku, and GPT-5 Mini. Because the resulting multi-label matrix was highly sparse, with approximately 98% of report–label pairs corresponding to shared negatives, raw percentage agreement would substantially overestimate concordance. We therefore evaluated chance-corrected agreement using Cohen’s κ for pairwise model comparisons and Fleiss’ κ for agreement across the complete seven-model panel. Mean pairwise Cohen’s κ was approximately 0.83 at the individual-label level and increased to 0.87 after related labels were collapsed into parent finding groups, indicating that a substantial component of apparent disagreement resulted from differences in label granularity. Consistently, parent- group Fleiss’ κ showed strong seven-model concordance across the clinical finding groups (κ = 0.79–0.98), whereas other_findings showed lower agreement (κ = 0.56), reflecting its more heterogeneous and open-ended vocabulary (Figure 5). 18Prabhu et al. Fig. 5: Inter-model concordance of LLM-derived labels. Left: Distribution of votes across seven models for active report–tag pairs, where the vote denotes the number of models (1–7) identifying a finding as present; unanimous-negative pairs (0/7) are excluded. Middle: Fleiss’ κ by parent finding group, showing strong agreement across clinical categories. Right: Model-specific 6-vs-1 dis- agreements, distinguishing lone misses from lone positive calls. Residual disagreement was dominated by synonym routing, whereby mod- els identified the same underlying finding but assigned different keys. Common examples included effusion versus pleural_effusion and the nodule fam- ily (nodules, solitary_nodules, and solitary_pulmonary_nodule). When solitary_pulmonary_nodule was missed under its specific key, nodules was assigned in 96% of such cases; similarly, pleural_effusion was assigned in 95% of cases in which effusion was missed. The open-ended other_findings field further contributed to this naming variability, reinforcing the need for canonical- ization. One notable exception was artifacts, which exhibited model-specific over-extraction rather than synonym substitution and consequently behaved as a source of annotation noise. For large-scale downstream label expansion, we selected Gemini 3 Flash for its inclusive extraction behavior relative to the other models. This stage is sepa- rate from Qwen3-based construction of the internal pretraining targets. General and specific labels may both be emitted and are examined during canonicaliza- tion. Without expert reference labels, this comparison does not establish clinical precision or recall. Inter-model agreement measures reproducibility of report extraction, and shared model errors can remain undetected. Parent-group agree- ment may also conceal errors in fine-grained distinctions. The expanded labels should therefore be interpreted as report-derived weak supervision rather than independently validated image-level ground truth. Supplementary Section S1.2 details parsing, normalization, and concordance analysis. 4.3 Implementation Pretraining MedCLIP is pretrained using AdamW (lr=2e-5) with a one-cycle scheduler. The text encoder is BiomedVLP-CXR-BERT, and the image encoder is InternViT (embedding size=512). For training the VLM, we use a learning rate of 1e-5 with AdamW and a warmup Med-AR for Long-Tailed Chest X-Ray Classification19 ratio of 0.05. Fine-Tuning ML-Decoder with k=N (no. of classes), where k is size of group queries serves as the classification head, processing the final hidden state of the image encoder (embedding size = 1024). The model is fine-tuned using Asym- metric Loss, AdamW, and a OneCycle scheduler (batch size = 2304, max lr = 5e-05). We used 8XH100 GPUs for both MedCLIP and Med-AR training. 4.4 Evaluation Metrics and Statistical Analysis Evaluation metrics. We evaluate downstream performance using AUROC, AUPRC, sensitivity, specificity, and Excess Area Under the Risk–Coverage Curve (EAURC). Since the task is multi-label classification, each abnormality is treated as a one-vs-rest binary prediction problem, and metrics are computed per label before being averaged over all labels as well as over the head, medium, and tail subsets. We report AUROC as a threshold-independent measure of ranking quality, and AUPRC as a complementary metric that is especially informative under class imbalance, where positive cases may be rare [37]. To characterize thresholded op- erating performance, we also report sensitivity and specificity, which quantify the ability to recover positive cases while avoiding false alarms. For public-dataset operating-point results, thresholds are selected per label by maximizing Youden’s J on the validation set, as described in Supplementary Section S2.1. To assess uncertainty-aware reliability, we use EAURC, a selective-prediction metric that evaluates how well model confidence ranks predictions by difficulty [14]. Lower EAURC indicates a lower excess selective risk under the evaluated confidence ordering. This metric does not measure probability calibration or establish clinical deployment readiness; in sparse multi-label tasks, ordinary error risk can be dominated by negative examples. Statistical significance tests. We use two complementary statistical tests to as- sess whether differences between pretraining paradigms are systematic. First, we apply the Wilcoxon signed-rank test to paired per-label AUROC and AUPRC values [46]. This test is appropriate because the comparison is paired at the label level, and the distribution of per-label metric differences is not guaranteed to be Gaussian. It therefore provides a robust non-parametric assessment of whether one pretraining strategy tends to outperform the other across all labels, or within the head and tail subsets. Second, we use DeLong’s test to compare label-wise AUROCs between mod- els [8]. DeLong’s test is designed for correlated ROC curves and is well suited here because the competing models are evaluated on the same test cases for each label. Unlike the Wilcoxon test, which summarizes paired differences across la- bels, DeLong’s test identifies whether the AUROC gap for a specific abnormality is statistically significant. The reported significance comparisons are interpreted as nominal, ex- ploratory evidence: correlated labels, multiple comparisons, repeated studies, and training variability limit inference. Non-significance does not establish 20Prabhu et al. equivalence, and smaller p-values do not measure larger effect sizes. Together, these tests provide complementary evidence: the Wilcoxon signed-rank test assesses whether one pretraining paradigm is consistently better across labels, while DeLong’s test identifies which individual abnormalities show significant AUROC differences. 5 Results and Discussion 5.1 Main quantitative comparison Table 3 summarizes mean AUROC and mean AUPRC across overall, head, medium, and tail label groups for the four evaluation benchmarks. A clear pat- tern emerges on the three public datasets, although the preferred autoregressive variant differs across them. On PadChest, autoregressive pretraining consistently outperforms Med-CLIP across all prevalence groups, with Med-AR-2B achieving the best results through- out. Specifically, mean AUROC improves from 0.8657 to 0.8800 overall, from 0.8425 to 0.8571 for head labels, from 0.8531 to 0.8674 for medium labels, and from 0.9002 to 0.9143 for tail labels. The corresponding AUPRC values also in- crease consistently, from 0.4310 to 0.4595 overall, from 0.4428 to 0.4706 for head labels, from 0.3678 to 0.3941 for medium labels, and from 0.4762 to 0.5072 for tail labels. A similar but even stronger trend is observed on MIMIC-CXR, where Med- AR-8B is the best-performing model across all groups. Overall mean AUROC in- creases from 0.8777 for Med-CLIP to 0.8971 for Med-AR-8B, while overall mean AUPRC rises from 0.2315 to 0.2769. The gains are visible across all prevalence regimes: for head labels, AUROC improves from 0.8756 to 0.8918 and AUPRC from 0.4010 to 0.4432; for medium labels, AUROC improves from 0.8841 to 0.9028 and AUPRC from 0.1902 to 0.2435; and for tail labels, AUROC improves from 0.8733 to 0.8967 and AUPRC from 0.1033 to 0.1441. CheXpert provides a complementary test of this trend and highlights a stronger role for decoder scale. Med-AR-8B achieves the best AUROC and AUPRC in every prevalence group, lifting overall AUROC from 0.8585 (Med- CLIP) to 0.8759 and overall AUPRC from 0.2142 to 0.2391, with consistent gains for head labels (0.8433 to 0.8565 AUROC; 0.3710 to 0.3978 AUPRC), medium labels (0.8790 to 0.8920 AUROC; 0.1802 to 0.2054 AUPRC), and tail labels (0.8526 to 0.8788 AUROC; 0.0922 to 0.1151 AUPRC). In contrast, Med-AR-2B does not surpass Med-CLIP on this benchmark and is lowest across every preva- lence group for both AUROC and AUPRC. Autoregressive supervision therefore remains beneficial on CheXpert, but the advantage is concentrated in the larger decoder, and a smaller autoregressive model is insufficient to outperform a strong contrastive baseline on this dataset. As shown in Section 5.2, paired per-label Wilcoxon results support most of the Med-AR-8B improvements on the public datasets at the nominal 0.05 level. The exception is PadChest medium-label AUPRC (p = 0.0645). These tests Med-AR for Long-Tailed Chest X-Ray Classification21 Table 3: Mean AUROC and AUPRC for overall, head, medium, and tail tags across datasets. The best value within each metric block is shown in bold, and the second-best value is underlined. Dataset Group mAUROCmAUPRC Med-CLIP Med-AR-8B (ours) Med-AR-2B (ours) Med-CLIP Med-AR-8B (ours) Med-AR-2B (ours) Internal Overall0.90810.91110.91350.28620.27540.2842 Head0.89470.89830.90080.42890.43670.4415 Medium0.90690.9100 0.91240.24340.22920.2452 Tail0.92270.92530.92740.18970.16390.1689 PadChest Overall0.86570.87220.88000.43100.44690.4595 Head0.84250.84970.85710.44280.45830.4706 Medium0.85310.8595 0.86740.36780.38180.3941 Tail0.90020.90600.91430.47620.49400.5072 MIMIC Overall0.87770.89710.88590.23150.27690.2541 Head0.87560.89180.88460.40100.44320.4264 Medium0.88410.90280.88810.19020.24350.2175 Tail0.87330.89670.88510.10330.14410.1185 CheXpert Overall0.85850.87590.84220.21420.23910.2039 Head0.84330.85650.83100.37100.39780.3614 Medium0.87900.89200.85860.18020.20540.1682 Tail0.8526 0.87880.83650.09220.11510.0829 describe the observed label-wise differences and do not isolate the contribution of the pretraining objective. The internal benchmark presents a more nuanced picture and is best de- scribed as broadly comparable between the two paradigms, with a modest edge for autoregressive pretraining in AUROC but a mixed pattern in AUPRC. In terms of AUROC, Med-AR-2B performs best across all groups, improving from 0.9081 to 0.9135 overall, from 0.8947 to 0.9008 for head labels, from 0.9069 to 0.9124 for medium labels, and from 0.9227 to 0.9274 for tail labels. However, the AUPRC results are less uniform. Med-AR-2B achieves the best performance for the head and medium groups, increasing AUPRC from 0.4289 to 0.4415 and from 0.2434 to 0.2452, respectively, whereas Med-CLIP remains strongest over- all and on the tail subset, with AUPRC values of 0.2862 overall and 0.1897 on tail labels. Consistent with this mixed aggregate pattern, the statistical analysis in subsection 5.2 shows that several internal comparisons do not reach signifi- cance, especially for AR-8B and for the tail subset, supporting the interpretation that Med-AR and Med-CLIP are broadly comparable on the internal benchmark rather than cleanly separated. The preferred autoregressive variant is dataset-dependent. Med-AR-2B achieves the highest AUROC on Internal and the best discrimination on PadChest, whereas Med-AR-8B is strongest on MIMIC-CXR and CheXpert. These results indicate that decoder scale interacts with the training recipe and downstream dataset; they do not establish the objective itself as the primary source of improvement. 22Prabhu et al. 5.2 Statistical significance across labels To determine whether the aggregate improvements in Table 3 are systematic across labels rather than driven by a small subset of abnormalities, we applied the Wilcoxon signed-rank test to paired per-label AUROC and AUPRC values. Table 4 reports results for the overall, head, medium, and tail groups, comparing both Med-AR-8B and Med-AR-2B against Med-CLIP. The public benchmarks generally support the performance trends in Table 3. On PadChest, both Med-AR variants have nominally significant overall gains in AUROC and AUPRC. Med-AR-2B also has nominally significant gains in every subgroup; Med-AR-8B does so except for medium-label AUPRC (p = 0.0645). On MIMIC-CXR, both variants have nominally significant AUROC and AUPRC gains in all groups. Effect magnitudes should be read from the metric differences, rather than inferred from the relative sizes of the p-values. CheXpert shows the most asymmetric pattern between the two AR variants. Med-AR-8B achieves highly significant improvements over Med-CLIP across the overall benchmark and all prevalence groups, with AUROC p-values ranging from 2.17e−15 overall to 3.01e−06 on the tail subset, and AUPRC p-values from 1.08e−12 overall to 1.27e−03 on the tail subset. In contrast, Med-AR- 2B fails to significantly outperform Med-CLIP on CheXpert: p-values are near 1.00 for all overall, head, and medium comparisons, and no significant AUPRC gain is observed on any subset. This directly corroborates the aggregate finding that Med-AR-2B does not surpass Med-CLIP on CheXpert and shows that the behavior is systematic at the per-label level rather than a metric-averaging ar- tifact. Taken together with PadChest and MIMIC-CXR, these results indicate that the overall advantage of autoregressive pretraining on public benchmarks is broad and label-consistent, but depends critically on decoder scale: Med-AR- 8B has nominally significant overall gains on all three public datasets, whereas Med-AR-2B has such gains on PadChest and MIMIC-CXR but not CheXpert. The internal dataset presents a more nuanced picture, consistent with the broadly comparable aggregate results reported in the previous subsection. For Med-AR-8B, none of the Wilcoxon comparisons reach significance, indicating that its small aggregate improvements over Med-CLIP are not consistently ex- pressed at the per-label level. In contrast, Med-AR-2B shows statistically signif- icant AUROC gains for the overall, head, and medium groups, and also achieves a significant AUPRC improvement on the head subset. However, neither au- toregressive variant shows significant gains on the internal tail subset, and most AUPRC comparisons remain non-significant. This pattern supports a balanced interpretation of the internal benchmark: Med-AR remains competitive and of- ten favorable, particularly for AUROC and for the head and medium groups, but the differences relative to Med-CLIP are not uniformly strong enough to support the same clear superiority observed on the public datasets. Med-AR for Long-Tailed Chest X-Ray Classification23 (a) Internal – Head (nodule) (b) Internal – Medium (granuloma) (c) Internal – Tail (medi- astinal mass) (d) PadChest – Head (infil- trates) (e) PadChest – Medium (nodule) (f) PadChest – Tail (calci- fied granuloma) (g) MIMIC-CXR – Head (nodules) (h) MIMIC-CXR – Medium (granuloma) (i) MIMIC-CXR – Tail (me- diastinal mass) (j) CheXpert – Head (nod- ules) (k) CheXpert – Medium (granuloma) (l) CheXpert – Tail (medi- astinal mass) Fig. 6: ROC curves comparing Med-CLIP, Med-AR-8B, and Med-AR-2B across four datasets. Rows correspond to Internal, PadChest, MIMIC-CXR, and CheX- pert, while columns correspond to representative head-, medium-, and tail- prevalence abnormalities, respectively. 24Prabhu et al. Table 4: Wilcoxon signed-rank test results comparing autoregressive and CLIP- based pretraining using per-label AUROC and AUPRC across datasets for all, head, medium, and tail tags. Dataset Tag Group Model AUROC Stat. AUROC p-val AUPRC Stat. AUPRC p-val Internal All AR-8B352.00.1578533.00.9080 AR-2B215.00.0023495.00.7982 Head AR-8B27.00.059439.00.2131 AR-2B17.00.012325.00.0453 Medium AR-8B31.00.169864.00.9045 AR-2B21.00.047157.00.7928 Tail AR-8B57.00.619682.00.9710 AR-2B37.00.178890.00.9933 PadChest All AR-8B48.04.36e-0543.02.35e-05 AR-2B1.03.73e-094.01.30e-08 Head AR-8B6.00.01379.00.0322 AR-2B0.09.77e-040.09.77e-04 Medium AR-8B3.00.00989.00.0645 AR-2B0.00.001950.00.00195 Tail AR-8B7.00.01861.00.00195 AR-2B1.00.001951.00.00195 MIMIC All AR-8B206.01.04e-17172.04.33e-18 AR-2B1032.08.42e-101164.08.90e-09 Head AR-8B0.01.46e-111.02.91e-11 AR-2B13.01.28e-0960.01.43e-06 Medium AR-8B10.01.56e-1016.06.15e-10 AR-2B189.00.0038158.07.75e-04 Tail AR-8B53.06.49e-0743.01.88e-07 AR-2B127.04.08e-04185.00.00959 CheXpert All AR-8B524.02.17e-15818.01.08e-12 AR-2B5418.01.00e+00 4175.00.9905 Head AR-8B3.01.82e-1131.08.64e-09 AR-2B643.01.00e+00555.00.9969 Medium AR-8B72.06.41e-0792.04.08e-06 AR-2B671.01.00e+00477.00.8876 Tail AR-8B81.03.01e-06167.00.00127 AR-2B549.00.9958408.00.7072 Overall, the Wilcoxon analysis reinforces the central narrative of the pa- per. On PadChest, MIMIC-CXR, and CheXpert, autoregressive pretraining— most consistently through Med-AR-8B—has nominally significant improvements over Med-CLIP in most prevalence-group comparisons, whereas on the internal benchmark the two paradigms are better characterized as broadly comparable, with the strongest support for Med-AR concentrated in Med-AR-2B and in the higher-prevalence groups. 5.3 Label-wise AUROC differences We further examined class-wise behavior using DeLong’s test for paired AU- ROCs. Figures 7 and 8 visualize the AUROC difference between autoregres- sive pretraining and Med-CLIP for each abnormality, separately for the head, Med-AR for Long-Tailed Chest X-Ray Classification25 medium, and tail subsets in the Internal, MIMIC-CXR, PadChest, and CheX- pert benchmarks. Statistically significant differences under DeLong’s test are marked in the figures. The public datasets again show a clear advantage for autoregressive pre- training. On MIMIC-CXR, the majority of labels across the head, medium, and tail groups show positive AUROC differences, and many of these gains are sta- tistically significant. The largest improvements are particularly evident among several medium and tail abnormalities, supporting the view that autoregres- sive supervision is especially beneficial when clinically relevant findings are less frequent or semantically more specific. PadChest exhibits a similarly favorable pattern, with consistently positive shifts across all three prevalence groups and multiple significant improvements among both common and infrequent abnor- malities. CheXpert follows the same trend when compared against Med-AR-8B: positive AUROC differences dominate the head, medium, and tail subsets, and a substantial fraction of labels reach statistical significance, including several rare abnormalities. This label-level evidence complements the Wilcoxon analysis and confirms that the Med-AR-8B gains on CheXpert are distributed across many abnormalities rather than concentrated in a handful of high-prevalence classes. Consistent with the earlier aggregate and Wilcoxon findings, Med-AR-2B on CheXpert does not produce a comparable distribution of positive significant differences, again indicating that the CheXpert benefit is tied to decoder scale. The internal dataset is more heterogeneous. In the head and medium groups, the class-wise AUROC differences mostly favor autoregressive pretraining, con- sistent with the Wilcoxon results for AR-2B. However, the internal tail subset remains mixed, with some abnormalities benefiting from AR and others contin- uing to favor Med-CLIP. This label-wise heterogeneity explains why the internal benchmark, despite favorable aggregate metrics for AR, yields more limited sta- tistical support at the tail level. Overall, the DeLong analysis reinforces the central pattern of the paper: autoregressive pretraining produces broad and con- sistent gains on the public datasets, while the internal benchmark shows a more selective advantage concentrated in higher-prevalence groups and clinically im- portant individual findings. 5.4 Uncertainty-aware evaluation Table 5 reports Excess Area Under the Risk–Coverage Curve (EAURC), where lower values indicate better uncertainty ranking and, therefore, more effective selective prediction. This analysis complements the discrimination results by testing whether model confidence is informative about prediction correctness, rather than only whether the predicted scores separate positives from negatives. The public datasets again favor autoregressive pretraining, although the best- performing variant varies across them. On MIMIC-CXR, Med-AR-8B achieves the lowest EAURC across all groups, reducing the overall value from 0.196 for Med-CLIP to 0.169, with consistent improvements for head labels (0.365 to 0.334), medium labels (0.186 to 0.146), and tail labels (0.038 to 0.027). Med-AR- 2B also improves over Med-CLIP in every group, although it remains slightly 26Prabhu et al. (a) Internal (b) MIMIC-CXR Fig. 7: Differences in AUROC, ∆, between autoregressive and Med-CLIP pre- training for (a) Internal and (b) MIMIC-CXR. Statistically significant differ- ences under the DeLong test are marked with (*). weaker than Med-AR-8B. A similar trend is observed on PadChest, where Med- AR-8B again provides the strongest uncertainty ranking, reducing EAURC from 0.314 to 0.217 overall, from 0.397 to 0.306 for head labels, from 0.351 to 0.240 for medium labels, and from 0.198 to 0.107 for tail labels. Med-AR-2B is consistently second-best on both datasets. The CheXpert results add an important and somewhat unexpected nuance. Both autoregressive variants outperform Med-CLIP on EAURC across all preva- lence groups, but here Med-AR-2B, not Med-AR-8B, achieves the best un- certainty ranking. Specifically, Med-AR-2B reduces EAURC from 0.212 (Med- CLIP) to 0.162 overall, from 0.406 to 0.356 for head labels, from 0.190 to 0.113 for medium labels, and from 0.040 to 0.018 for tail labels, with Med-AR-8B second- best throughout. This is particularly notable because Med-AR-2B has lower AUROC and AUPRC than Med-CLIP on CheXpert. In other words, stronger discrimination does not automatically translate into better uncertainty rank- Med-AR for Long-Tailed Chest X-Ray Classification27 (a) PadChest (b) CheXpert Fig. 8: Differences in AUROC, ∆, between autoregressive and Med-CLIP pre- training for (a) PadChest and (b) CheXpert. Statistically significant differences under the DeLong test are marked with (*). ing: despite being a weaker classifier on CheXpert, Med-AR-2B produces confi- dence scores that rank its own errors more effectively than either Med-CLIP or Med-AR-8B. This dissociation between discrimination and selective-prediction behavior highlights the value of evaluating these two axes independently when comparing pretraining paradigms. In contrast, the internal benchmark shows the opposite pattern. Here, Med- CLIP achieves the lowest EAURC overall and within each subgroup, with values of 0.014 overall, 0.036 for head labels, 0.004 for medium labels, and 0.002 for tail labels. Both Med-AR variants perform worse on this benchmark, with Med-AR- 8B ahead of Med-AR-2B overall and for head labels, and Med-AR-2B ahead for medium and tail labels. Both remain above Med-CLIP in every group. This is consistent with the more mixed behavior already observed on the internal dataset in the previous subsections: although Med-AR remains competitive in discrim- 28Prabhu et al. ination metrics, those gains do not translate into better confidence ordering in this more heterogeneous setting. Taken together, the uncertainty-aware evaluation reinforces the broader nar- rative of the paper while adding two important qualifications. First, on all three public benchmarks, Med-AR-8B improves AUROC, AUPRC, and EAURC rela- tive to Med-CLIP under the evaluated protocol; however, the optimal AR variant for uncertainty ranking is not always the one that maximizes discrimination, as clearly demonstrated on CheXpert. Second, on the internal dataset Med-CLIP retains a clear advantage in EAURC, suggesting that stronger representation transfer and better uncertainty ordering need not always coincide. These obser- vations make uncertainty-aware evaluation an essential complementary axis for comparing pretraining paradigms in clinical imaging. Table 5: Excess Area Under the Risk–Coverage Curve (EAURC) for Med-CLIP, Med-AR-8B, and Med-AR-2B across datasets. Values are averaged over all la- bels (Overall) and separately over head, medium, and tail labels. Lower values indicate better uncertainty ranking. The best value in each row is shown in bold, and the second-best value is underlined. DatasetGroup Med-CLIP EAURC Med-AR-8B (ours) EAURC Med-AR-2B (ours) EAURC Internal Overall0.0140.0700.073 Head0.0360.1730.194 Medium0.0040.0180.016 Tail0.0020.0220.013 MIMIC Overall0.1960.1690.179 Head0.3650.3340.345 Medium0.1860.1460.165 Tail0.0380.0270.028 PadChest Overall0.3140.2170.274 Head0.3970.3060.358 Medium0.3510.2400.315 Tail0.1980.1070.153 CheXpert Overall0.2120.1910.162 Head0.4060.3840.356 Medium0.1900.1590.113 Tail0.0400.0310.018 Med-AR for Long-Tailed Chest X-Ray Classification29 5.5 Comparison of pretrained encoders The uncertainty-aware evaluation above compares Med-AR directly against Med-CLIP. We next broaden that comparison to ask whether the observed gains persist when the image encoder is initialized from a wider set of rep- resentation learning strategies, while keeping the downstream classifier fixed. To this end, we replace the pretrained encoder with alternatives spanning self-supervised, supervised, biomedical contrastive, knowledge-enhanced, and radiology-specific vision–language pretraining, and fine-tune all of them under the same ML-Decoder downstream setup on CheXpert, MIMIC-CXR, and PadChest. Specifically, we compare RAD-DINO, an image-only self-supervised medi- cal encoder trained beyond text supervision [32]; ARK (Ark + ), a supervised chest radiography foundation model trained from heterogeneous expert-labeled datasets without manual label consolidation [28,27,29]; BiomedCLIP, a large- scale biomedical contrastive model pretrained on scientific image–text pairs [53]; MedKLIP, a knowledge-enhanced radiology vision–language pretraining method that aligns medical entities with image patches [47]; the EVA-based encoder (EVA-Base in the tables); BioViL-T, a radiology-specific contrastive vision– language model from the BiomedVLP family that builds on the CXR-BERT text encoder and improved domain-specific textual semantics [3]; CheXFound, a self-supervised chest X-ray foundation model pretrained on a large curated CXR corpus and originally paired with a global–local representation integration strategy for downstream adaptation [50]; and our proposed Med-AR-2B and Med-AR-8B encoders. Table 6 summarizes the resulting performance across the three public benchmarks. Label-wise results are organized by dataset in Supple- mentary Sections S3–S6. Table 6 shows that the autoregressive encoders remain the strongest rep- resentation learners under a fixed downstream setup, and this conclusion now holds across all three public datasets. On CheXpert, Med-AR-8B achieves the best mAUROC and mAUPRC for the head, medium, and tail subsets, with MedCLIP consistently second-best and ARK/CheXFound close behind on the head subset. On MIMIC-CXR, Med-AR-8B is again best across all prevalence groups, while Med-AR-2B is second-best throughout, clearly ahead of the non- autoregressive baselines. On PadChest, Med-AR-2B is the strongest model across head, medium, and tail subsets, with Med-AR-8B second-best and ARK/ CheX- Found leading among the non-autoregressive alternatives. Among the non-AR baselines, ARK and CheXFound are the most competitive on the head and tail groups, whereas the other radiology-specific and biomedical baselines, including MedKLIP, EVA-Base, and BioViL-T, remain clearly below the best Med-AR variants on both ranking-based metrics for every dataset. 30Prabhu et al. Table 6: Comparison of vision encoders initialized from different methods across CheXpert, MIMIC-CXR, and PadChest using mean AUROC and mean AUPRC. All encoders are fine-tuned with the same ML-Decoder head. Within each dataset, metric, and prevalence group, the best result is shown in bold and the second-best is underlined. H=head, M=medium, and T=tail. MethodGrp CheXpertMIMIC-CXRPadChest mAUROC mAUPRCmAUROC mAUPRCmAUROC mAUPRC EVA-Base H0.79810.29180.82500.30130.79610.3463 M0.82230.09960.81150.08040.79340.2480 T0.78430.03330.79520.03260.81850.2581 RAD-DINO H 0.76320.24730.84400.32870.77770.3109 M0.78470.06140.83800.10320.77640.2224 T 0.74600.01930.82890.03970.80560.2406 ARK H0.84300.36800.88370.42020.84190.4384 M0.85930.14510.88090.18170.85100.3633 T0.83250.06490.88160.11110.90470.4747 CheXFound H0.84290.36920.88090.40980.84870.4558 M0.86700.16650.88490.19870.85840.3770 T0.83840.08150.87360.10080.89500.4476 BiomedCLIP H0.73920.23600.85820.37190.66300.2034 M0.76040.07150.85360.15420.67050.1471 T0.73110.02690.84250.07910.72410.1970 MedKLIP H0.78470.27860.81120.28390.81320.3812 M0.79800.07770.78660.06210.81720.3045 T0.74110.02010.76280.02980.83440.3406 BioViL-T H 0.66040.16760.71170.18810.66260.1858 M 0.62170.02000.64240.02070.66320.1510 T0.61130.00720.58330.00610.71040.2087 MedCLIP H0.84330.37100.87560.40100.84250.4428 M0.87900.18020.88410.19020.85310.3678 T0.85260.09220.87330.10330.90020.4762 Med-AR-2B (ours) H 0.83100.36140.88460.42640.85710.4706 M 0.85860.16820.88810.21750.86740.3941 T0.83650.08290.88510.11850.91430.5072 Med-AR-8B (ours) H0.85650.39780.89180.44320.84970.4583 M0.89200.20540.90280.24350.85950.3818 T0.87880.11510.89670.14410.90600.4940 Med-AR for Long-Tailed Chest X-Ray Classification31 This comparison extends the analysis beyond a single contrastive baseline un- der a common downstream head. It is not a component ablation: the pretrained systems differ in architecture, pretraining data, and optimization. The results support transfer performance under the evaluated setup, without attributing differences solely to the pretraining objective or excluding interactions with the downstream head. 5.6 Discussion Taken together, the results support five main observations. First, on all three public benchmarks—PadChest, MIMIC-CXR, and CheXpert—autoregressive pretraining improves mAUROC and mAUPRC over Med-CLIP with Med-AR-8B across all prevalence groups, with nominal statis- tical support in most comparisons, with the strongest and most consistent gains delivered by Med-AR-8B. On MIMIC-CXR and CheXpert, Med-AR-8B is best overall and in every prevalence subgroup; on PadChest, the smaller Med-AR-2B is best while Med-AR-8B remains a strong second-best. The consistency of this pattern across aggregate metrics, paired Wilcoxon tests, and label-wise DeLong comparisons suggests a broad improvement in representation transfer rather than an artifact of a few abnormalities. The tail-label gains are consistent with the hypothesis that structured and region-aware autoregressive supervision benefits rare-finding recognition, but do not directly demonstrate the proposed visual-grounding mechanism. Second, the internal benchmark is more nuanced. Med-AR-2B shows a modest mAUROC edge with partial statistical support, but mAUPRC results are mixed, tail-label separation is not uniform, and Med-CLIP retains a clear EAURC advantage in every subgroup. The two paradigms are therefore best described as broadly comparable on internal data, suggesting that the benefit of autoregressive pretraining is partly conditioned by dataset properties such as reporting consistency, label construction quality, and institutional heterogeneity. Third, the EAURC analysis reveals that stronger discrimination does not guarantee better confidence-based risk ordering. Med-AR improves EAURC on all three public benchmarks, but CheXpert provides a striking example where discrimination and selective-prediction behavior decouple: Med-AR-2B has lower AUROC and AUPRC than Med-CLIP yet achieves the best EAURC across every CheXpert subgroup. On the internal dataset, the opposite dissociation occurs, with Med-CLIP achieving the best EAURC despite Med-AR matching or exceeding it on discrimination metrics. Together, these findings underscore the importance of evaluating selective-prediction behavior as a distinct axis rather than assuming it follows discrimination gains. Fourth, the encoder comparison across CheXpert, MIMIC-CXR, and Pad- Chest shows that Med-AR encoders outperform not only Med-CLIP but also a broad set of self-supervised, supervised, biomedical contrastive, knowledge- enhanced, and radiology-specific VLP baselines under the same ML-Decoder head, supporting the effectiveness of the complete pretraining recipe under this downstream setup. 32Prabhu et al. Finally, the preferred autoregressive variant is dataset-dependent: Med-AR- 2B leads on Internal and PadChest, while Med-AR-8B leads on MIMIC-CXR and CheXpert. This is most consequential on CheXpert, where Med-AR-2B fails to outperform Med-CLIP on discrimination while Med-AR-8B does so decisively. This suggests an interaction between decoder capacity, the pretraining recipe, and the downstream task. The present experiments do not isolate the objective as the primary driver of the gains. Overall, these findings identify radiology-native autoregressive pretraining as a strong foundation for long-tailed CXR classification on public benchmarks, while the mixed internal results highlight that representation superiority is not absolute and depends on both dataset characteristics and the evaluation axis under consideration. 5.7 Limitations The comparison controls the downstream architecture, but the pretraining recipes differ in supervision and may differ in initialization and compute. No matched-supervision component ablation is available, so the results can- not establish a causal advantage of next-token prediction alone. The additional pretrained-encoder comparison also does not equalize architecture or pretraining data. The expanded MIMIC-CXR and CheXpert labels are derived from reports. Inter-model concordance does not establish expert-level correctness, and unmen- tioned findings may be encoded as negatives. Generated reasoning and promi- nence targets may introduce additional noise. The 1,000-report concordance sam- ple provides limited evidence for very rare labels, and the retained PadChest vocabulary excludes many less frequent findings. Each dataset is fine-tuned sep- arately; these experiments do not constitute deployment on an unseen institution without adaptation. The statistical analyses do not quantify training-seed variability or establish equivalence for non-significant comparisons. Label dependencies and multiple testing warrant caution in interpreting nominal significance. EAURC assesses selective prediction under the evaluated protocol rather than calibration or clin- ical utility. Independent expert validation, fuller reporting of data provenance and split construction, and prospective evaluation remain necessary to assess clinical applicability. 6 Conclusion and Future Work We investigated whether the choice of vision–language pretraining recipe— contrastive versus autoregressive—affects downstream transfer for long-tailed multi-label chest X-ray classification. To reduce downstream architectural con- founding, we compared MedCLIP-style contrastive pretraining and InternVL3- based autoregressive pretraining (Med-AR-2B and Med-AR-8B) using a shared InternViT vision architecture and a common ML-Decoder classification head, Med-AR for Long-Tailed Chest X-Ray Classification33 and evaluated all models on an internal benchmark together with three public datasets: PadChest, MIMIC-CXR, and CheXpert. Across the three public benchmarks, Med-AR-8B improves downstream AU- ROC and AUPRC over Med-CLIP in each prevalence group, with nominally significant paired per-label gains in most comparisons. Med-AR-8B performs best on MIMIC-CXR and CheXpert, while Med-AR-2B performs best on Pad- Chest. The encoder comparison also favors Med-AR under the evaluated down- stream setup. These observations support the complete autoregressive pretrain- ing recipe, while the separate contributions of objective, supervision, and decoder scale remain unresolved. On the internal benchmark, the two paradigms are better characterized as broadly comparable: Med-AR-2B shows a modest AUROC edge with partial sta- tistical support, but AUPRC results are mixed and Med-CLIP retains a clear ad- vantage in selective-prediction behavior as measured by EAURC. More broadly, our uncertainty-aware analysis uncovers a consequential decoupling between dis- crimination and confidence ranking. On CheXpert, Med-AR-2B achieves the best EAURC across all prevalence groups despite lower AUROC and AUPRC than Med-CLIP, while on the internal dataset the opposite dissociation oc- curs. Stronger discrimination therefore does not guarantee better confidence- based risk ordering, and prevalence-aware, uncertainty-aware evaluation should be treated as a first-class axis—rather than a secondary metric—when compar- ing medical vision–language pretraining strategies. Several directions remain open. First, disentangling the contribution of the autoregressive objective itself from its richer supervision—structured multi-section report templates, abnormality-focused auxiliary outputs, and region-level annotations—will require component-wise ablations that we leave to future work. Second, dedicated strategies for rare-label learning, such as curriculum scheduling, meta-learning, and focused autoregressive pretraining on curated rare-finding corpora, may further close the gap on tail abnormalities and reduce residual variability across decoder scales. Third, integrating calibration techniques such as temperature scaling [26,48], conformal prediction [23,1], and post-hoc meta-models [39] would complement EAURC and support further assessment of reliability before clinical deployment. Finally, extending the LLM- based label-expansion framework to additional datasets and imaging modalities, and embedding the resulting confidence estimates into clinical workflows such as triage and radiologist–AI collaboration, are important next steps toward assessing real-world impact. In summary, radiology-native autoregressive pretraining provides a strong foundation for long-tailed chest X-ray classification on public benchmarks, par- ticularly for underrepresented findings. Its advantages, however, are not uniform across datasets or evaluation axes, and the Med-AR configuration that is best for discrimination need not coincide with the one that is best for selective prediction. These observations argue for a joint assessment of discrimination and uncertainty behavior when comparing medical vision–language pretraining paradigms, and 34Prabhu et al. position radiology-native Med-AR as a starting point for further research on reliable chest X-ray classification. 7 Statements and Declarations Author Contributions All authors contributed to the conception and design of the study. Janhavi Prabhu led the model training, methodology development, exper- imental design, and manuscript writing. Sahil contributed to model training, dataset preparation, label extraction and manuscript writing. Akshay Valsaraj contributed to implementation support and model development. Manoj Tadepalli and Preetham Putha provided oversight, technical guidance, and critical revisions of the manuscript. All authors reviewed and approved the final version of the manuscript. Data Availability We plan to release the derived MIMIC-CXR and CheXpert label annotations and split identifiers following publication, subject to the source datasets’ access and licensing requirements. Original images and reports remain subject to their respective access conditions. Disclosure of Interests. The authors have no competing interests to declare that are relevant to the content of this article. References 1. Angelopoulos, A., Bates, S., Malik, J., Jordan, M.I.: Uncertainty sets for image classifiers using conformal prediction (Sep 2020) 2. Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang, S., Tang, J., Zhong, H., Zhu, Y., Yang, M., Li, Z., Wan, J., Wang, P., Ding, W., Fu, Z., Xu, Y., Ye, J., Zhang, X., Xie, T., Cheng, Z., Zhang, H., Yang, Z., Xu, H., Lin, J.: Qwen2.5-VL technical report (Feb 2025) 3. Bannur, S., Hyland, S., Liu, Q., Perez-Garcia, F., Ilse, M., Castro, D.C., Boeck- ing, B., Sharma, H., Bouzid, K., Thieme, A., Schwaighofer, A., Wetscherek, M., Lungren, M.P., Nori, A., Alvarez-Valle, J., Oktay, O.: Learning to Exploit Tem- poral Structure for Biomedical Vision-Language Processing . In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). p. 15016– 15027. IEEE Computer Society, Los Alamitos, CA, USA (Jun 2023). https:// doi.org/10.1109/CVPR52729.2023.01442, https://doi.ieeecomputersociety. org/10.1109/CVPR52729.2023.01442 4. Boecking, B., Usuyama, N., Bannur, S., Castro, D.C., Schwaighofer, A., Hyland, S., Wetscherek, M., Naumann, T., Nori, A., Alvarez-Valle, J., et al.: Making the most of text semantics to improve biomedical vision–language processing. In: European conference on computer vision. p. 1–21. Springer (2022) 5. Bustos, A., Pertusa, A., Salinas, J.M., de la Iglesia-Vayá, M.: Padchest: A large chest x-ray image dataset with multi-label annotated reports. Medical image anal- ysis 66, 101797 (2019), https://api.semanticscholar.org/CorpusID:58981612 6. Chen, J., Yang, J., Wu, H., Li, D., Gao, J., Zhou, T., Xiao, B.: Florence-VL: En- hancing vision-language models with generative vision encoder and depth-breadth fusion (Dec 2024) Med-AR for Long-Tailed Chest X-Ray Classification35 7. Chen, Z., Wu, J., Wang, W., Su, W., Chen, G., Xing, S., Zhong, M., Zhang, Q., Zhu, X., Lu, L., et al.: Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. p. 24185–24198 (2024) 8. DeLong, E.R., DeLong, D.M., Clarke-Pearson, D.L.: Comparing the areas under two or more correlated receiver operating characteristic curves: A nonparametric approach. Biometrics 44(3), 837–845 (1988). https://doi.org/10.2307/2531595, https://pubmed.ncbi.nlm.nih.gov/3203132/ 9. Desai, K., Johnson, J.: Virtex: Learning visual representations from textual anno- tations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). p. 11162–11173 (June 2021) 10. El-Nouby, A., Klein, M., Zhai, S., Bautista, M.Á., Shankar, V., Toshev, A.T., Susskind, J.M., Joulin, A.: Scalable pre-training of large autoregressive image mod- els. In: Proceedings of the 41st International Conference on Machine Learning. Pro- ceedings of Machine Learning Research, vol. 235, p. 12371–12384. PMLR (2024) 11. Faghani, S., Moassefi, M., Rouzrokh, P., Khosravi, B., Baffour, F.I., Ringler, M.D., Erickson, B.J.: Quantifying uncertainty in deep learning of radiologic images. Ra- diology 308(2), e222217 (2023) 12. Fini, E., Shukor, M., Li, X., Dufter, P., Klein, M., Haldimann, D., Aitharaju, S., da Costa, V.G.T., Béthune, L., Gan, Z., Toshev, A., Eichner, M., Nabi, M., Yang, Y., Susskind, J., El-Nouby, A.: Multimodal autoregressive pre-training of large vision encoders. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). p. 9641–9654 (June 2025) 13. Geifman, Y., El-Yaniv, R.: Selectivenet: A deep neural network with an integrated reject option. In: International Conference on Machine Learning (2019), https: //api.semanticscholar.org/CorpusID:59316904 14. Geifman, Y., Uziel, G., El-Yaniv, R.: Bias-reduced uncertainty estimation for deep neural classifiers. In: International Conference on Learning Representations (2018), https://api.semanticscholar.org/CorpusID:52901777 15. Holste, G., Wang, S., Jiang, Z., Shen, T.C., Shih, G., Summers, R.M., Peng, Y., Wang, Z.: Long-tailed classification of thorax diseases on chest x-ray: A new bench- mark study. In: MICCAI Workshop on Data Augmentation, Labelling, and Imper- fections. p. 22–32. Springer (2022) 16. Holste, G., Zhou, Y., Wang, S., Jaiswal, A., Lin, M., Zhuge, S., Yang, Y., Kim, D., Nguyen-Mau, T.H., Tran, M.T., et al.: Towards long-tailed, multi-label disease classification from chest x-ray: Overview of the cxr-lt challenge. Medical Image Analysis p. 103224 (2024) 17. Huang, S.C., Shen, L., Lungren, M.P., Yeung, S.: GLoRIA: A multimodal global- local representation learning framework for label-efficient medical image recogni- tion. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE (Oct 2021) 18. Hüllermeier, E., Waegeman, W.: Aleatoric and epistemic uncertainty in machine learning: an introduction to concepts and methods. Mach. Learn. 110(3), 457–506 (Mar 2021) 19. Irvin, J., Rajpurkar, P., Ko, M., Yu, Y., Ciurea-Ilcus, S., Chute, C., Marklund, H., Haghgoo, B., Ball, R., Shpanskaya, K., Seekins, J., Mong, D.A., Halabi, S.S., Sandberg, J.K., Jones, R., Larson, D.B., Langlotz, C.P., Patel, B.N., Lungren, M.P., Ng, A.Y.: Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison (2019), https://arxiv.org/abs/1901.07031 36Prabhu et al. 20. Jing, D., He, X., Luo, Y., Fei, N., Yang, G., Wei, W., Zhao, H., Lu, Z.: Fineclip: Self-distilled region-based clip for better fine-grained understanding. In: Advances in Neural Information Processing Systems (2024) 21. Johnson, A., Pollard, T., Mark, R., Berkowitz, S., Horng, S.: MIMIC-CXR Database. PhysioNet (Jul 2024). https://doi.org/10.13026/4jqj-jw95, https: //doi.org/10.13026/4jqj-jw95, version 2.1.0 22. Johnson, A.E.W., Pollard, T., Mark, R., Berkowitz, S., Horng, S.: The mimic-cxr database (2019). https://doi.org/10.13026/C2JT1Q, https://physionet.org/ content/mimic-cxr/ 23. Karimi, H., Samavi, R.: Quantifying deep learning model uncertainty in conformal prediction. Toronto Metropolitan University (Jan 2024) 24. Kim, D.: Chexfusion: Effective fusion of multi-view features using transformers for long-tailed chest x-ray classification. 2023 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW) p. 2694–2702 (2023), https://api. semanticscholar.org/CorpusID:260704332 25. Kimi Team, Du, A., Yin, B., Xing, B., Qu, B., Wang, B., Chen, C., Zhang, C., Du, C., Wei, C., Wang, C., Zhang, D., Du, D., Wang, D., Yuan, E., Lu, E., Li, F., Sung, F., Wei, G., Lai, G., Zhu, H., Ding, H., Hu, H., Yang, H., Zhang, H., Wu, H., Yao, H., Lu, H., Wang, H., Gao, H., Zheng, H., Li, J., Su, J., Wang, J., Deng, J., Qiu, J., Xie, J., Wang, J., Liu, J., Yan, J., Ouyang, K., Chen, L., Sui, L., Yu, L., Dong, M., Dong, M., Xu, N., Cheng, P., Gu, Q., Zhou, R., Liu, S., Cao, S., Yu, T., Song, T., Bai, T., Song, W., He, W., Huang, W., Xu, W., Yuan, X., Yao, X., Wu, X., Li, X., Zu, X., Zhou, X., Wang, X., Charles, Y., Zhong, Y., Li, Y., Hu, Y., Chen, Y., Wang, Y., Liu, Y., Miao, Y., Qin, Y., Chen, Y., Bao, Y., Wang, Y., Kang, Y., Liu, Y., Dong, Y., Du, Y., Wu, Y., Wang, Y., Yan, Y., Zhou, Z., Li, Z., Jiang, Z., Zhang, Z., Yang, Z., Huang, Z., Huang, Z., Zhao, Z., Chen, Z., Lin, Z.: Kimi-VL technical report (Jun 2025) 26. Kull, M., Perello-Nieto, M., Kängsepp, M., Filho, T.S., Song, H., Flach, P.: Be- yond temperature scaling: Obtaining well-calibrated multiclass probabilities with dirichlet calibration (Oct 2019) 27. Ma, D., Pang, J., Gotway, M.B., Liang, J.: Foundation ark: Accruing and reusing knowledge for superior and robust performance. In: Medical Image Computing and Computer Assisted Intervention – MICCAI 2023. p. 651–662. Springer Nature Switzerland, Cham (2023) 28. Ma, D., Pang, J., Gotway, M.B., Liang, J.: A fully open AI foundation model applied to chest radiography. Nature p. 1–11 (2025) 29. Ma, D., Pang, J., Senthil Velan, S., Gotway, M.B., Liang, J.: Ark+: Super- vised training a single high-performance ai foundation model from many differ- ently labeled datasets—no label consolidation required. Medical Image Analy- sis 108, 103828 (2026). https://doi.org/10.1016/j.media.2025.103828, https: //w.sciencedirect.com/science/article/pii/S1361841525003743 30. Moon, J.H., Lee, H., Shin, W., Kim, Y.H., Choi, E.: Multi-modal understanding and generation for medical images and text via vision-language pre-training. IEEE Journal of Biomedical and Health Informatics 26(12), 6070–6080 (2022) 31. Park, W., Park, I., Kim, S., Ryu, J.B.: Robust asymmetric loss for multi-label long- tailed learning. 2023 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW) p. 2703–2712 (2023), https://api.semanticscholar.org/ CorpusID:260775967 32. Pérez-García, F., Sharma, H., Bond-Taylor, S., Bouzid, K., Salvatelli, V., Ilse, M., Bannur, S., Castro, D.C., Schwaighofer, A., Lungren, M.P., Wetscherek, M.T., Med-AR for Long-Tailed Chest X-Ray Classification37 Codella, N., Hyland, S.L., Alvarez-Valle, J., Oktay, O.: Exploring scalable medical image encoders beyond text supervision. Nature Machine Intelligence 7, 119–130 (2025). https://doi.org/10.1038/s42256-024-00965-w 33. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. p. 8748–8763. PmLR (2021) 34. Rao, Y., Zhao, W., Chen, G., Tang, Y., Zhu, Z., Huang, G., Zhou, J., Lu, J.: Dense- clip: Language-guided dense prediction with context-aware prompting. In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). p. 18082–18091 (June 2022) 35. Ridnik, T., Sharir, G., Ben-Cohen, A., Ben-Baruch, E., Noy, A.: Ml-decoder: Scalable and versatile classification head. 2023 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) p. 32–41 (2021), https://api. semanticscholar.org/CorpusID:244709117 36. Ridnik, T., Ben-Baruch, E., Zamir, N., Noy, A., Friedman, I., Protter, M., Zelnik- Manor, L.: Asymmetric loss for multi-label classification. In: Proceedings of the IEEE/CVF international conference on computer vision. p. 82–91 (2021) 37. Saito, T., Rehmsmeier, M.: The precision-recall plot is more informative than the roc plot when evaluating binary classifiers on imbalanced datasets. PLOS ONE 10(3), e0118432 (2015). https://doi.org/10.1371/journal.pone.0118432, https://pubmed.ncbi.nlm.nih.gov/25738806/ 38. Sensoy, M., Saleki, M., Julier, S., Aydogan, R., Reid, J.: Misclassification risk and uncertainty quantification in deep classifiers. In: 2021 IEEE Winter Conference on Applications of Computer Vision (WACV). vol. 86, p. 2483–2491. IEEE (Jan 2021) 39. Shen, M., Bu, Y., Sattigeri, P., Ghosh, S., Das, S., Wornell, G.: Post-hoc uncer- tainty learning using a dirichlet meta-model. Proc. Conf. AAAI Artif. Intell. 37(8), 9772–9781 (Jun 2023) 40. Sun, Q., Yu, Q., Cui, Y., Zhang, F., Zhang, X., Wang, Y., Gao, H., Liu, J., Huang, T., Wang, X.: Emu: Generative pretraining in multimodality. In: The Twelfth In- ternational Conference on Learning Representations (2024) 41. Wang, Z., Wu, Z., Agarwal, D., Sun, J.: Medclip: Contrastive learning from un- paired medical images and text. Proceedings of the Conference on Empirical Meth- ods in Natural Language Processing. Conference on Empirical Methods in Natural Language Processing 2022, 3876–3887 (2022), https://api.semanticscholar. org/CorpusID:252992913 42. Wang, Z., Yu, J., Yu, A.W., Dai, Z., Tsvetkov, Y., Cao, Y.: Simvlm: Sim- ple visual language model pretraining with weak supervision. arXiv preprint arXiv:2108.10904 (2021) 43. Wei, Y., Wang, X., Ong, H., Zhou, Y., Flanders, A., Shih, G., Peng, Y.: Enhancing disease detection in radiology reports through fine-tuning lightweight llm on weak labels. arXiv preprint arXiv:2409.16563 (2024) 44. Whata, A., Dibeco, K., Madzima, K., Obagbuwa, I.: Uncertainty quantification in multi-class image classification using chest x-ray images of covid-19 and pneumonia. Frontiers in Artificial Intelligence 7, 1410841 (2024) 45. Wiehe, A., Schneider, F., Blank, S., Wang, X., Zorn, H.P., Biemann, C.: Language over labels: Contrastive language supervision exceeds purely label-supervised clas- sification performance on chest x-rays. In: Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th 38Prabhu et al. International Joint Conference on Natural Language Processing: Student Research Workshop. p. 76–83. Association for Computational Linguistics, Stroudsburg, PA, USA (2022) 46. Wilcoxon, F.: Individual comparisons by ranking methods. Biometrics Bulletin 1(6), 80–83 (1945). https://doi.org/10.2307/3001968 47. Wu, C., Zhang, X., Zhang, Y., Wang, Y., Xie, W.: Medklip: Medical knowledge enhanced language-image pre-training for x-ray diagnosis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). p. 21315– 21326 (2023). https://doi.org/10.1109/ICCV51070.2023.01954 48. Xie, J., Chen, A.S., Lee, Y., Mitchell, E., Finn, C.: Calibrating language models with adaptive temperature scaling (Sep 2024) 49. Yang, A., Li, A., Yang, B., Zhang, B., Hui, B., Zheng, B., et al.: Qwen3 technical report. arXiv preprint arXiv:2505.09388 (2025), https://arxiv.org/abs/2505. 09388 50. Yang, Z., Xu, X., Zhang, J., Wang, G., Kalra, M.K., Yan, P.: Chest x-ray foundation model with global and local representations integration. IEEE Transactions on Medical Imaging 44(12), 4787–4799 (2025). https://doi.org/10.1109/TMI.2025. 3581907 51. You, K., Gu, J., Ham, J., Park, B., Kim, J., Hong, E.K., Baek, W., Roh, B.: Cxr- clip: Toward large scale chest x-ray language-image pre-training. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. p. 101–111. Springer (2023) 52. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: Coca: Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research (2022) 53. Zhang, S., Xu, Y., Usuyama, N., Xu, H., Bagga, J., Tinn, R., Preston, S., Rao, R., Wei, M., Valluri, N., Wong, C., Tupini, A., Wang, Y., Mazzola, M., Shukla, S., Liden, L., Gao, J., Lungren, M.P., Naumann, T., Wang, S., Poon, H.: Biomedclip: a multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs. arXiv preprint arXiv:2303.00915 (2023). https://doi.org/10. 48550/arXiv.2303.00915 54. Zhao, Y., Chen, S., Chen, Q., Hu, Z.: Combining loss reweighting and sample resampling for long-tailed instance segmentation. In: ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). p. 1–5 (2023). https://doi.org/10.1109/ICASSP49357.2023.10094303 55. Zhu, J., Wang, W., Chen, Z., Liu, Z., Ye, S., Gu, L., Tian, H., Duan, Y., Su, W., Shao, J., Gao, Z., Cui, E., Wang, X., Cao, Y., Liu, Y., Wei, X., Zhang, H., Wang, H., Xu, W., Li, H., Wang, J., Deng, N., Li, S., He, Y., Jiang, T., Luo, J., Wang, Y., He, C., Shi, B., Zhang, X., Shao, W., He, J., Xiong, Y., Qu, W., Sun, P., Jiao, P., Lv, H., Wu, L., Zhang, K., Deng, H., Ge, J., Chen, K., Wang, L., Dou, M., Lu, L., Zhu, X., Lu, T., Lin, D., Qiao, Y., Dai, J., Wang, W.: InternVL3: Exploring advanced training and test-time recipes for open-source multimodal models (Apr 2025) 56. Zou, K., Chen, Z., Yuan, X., Shen, X., Wang, M., Fu, H.: A review of uncertainty estimation and its application in medical imaging (Feb 2023) Supplementary Material Med-AR: Autoregressive Vision-Language Pretraining for Long-Tailed Chest X-Ray Classification and Uncertainty-Aware Evaluation This supplement provides label-construction details and the existing per-label results. Results are grouped by dataset so that discrimination and operating-point metrics can be inspected together. In the tables, ∗ denotes tail labels and † denotes medium-frequency labels; unmarked labels are head labels. Bold and underlined entries indicate the highest and second-highest distinct displayed values, respectively, with ties sharing rank. These marks are descriptive, not significance tests. Section Contents S1Target preparation and label concordance S2Cross-dataset summary and operating points S3Internal: consolidated per-label metrics S4CheXpert: all five per-label metrics S5MIMIC-CXR: all five per-label metrics S6PadChest: all five per-label metrics S1 Target Preparation and Label Concordance S1.1 Pretraining Target Structure The autoregressive targets use 15 anatomical subheadings with finding-specific fields, together with Findings, Impres- sion, Summary, Reasoning, and Prominence Score sections. Prominence scores range from 0 to 5 and are generated from report text. Qwen3 is used to construct these report-derived targets; expert bounding-box annotations provide separate region-level supervision for approximately 20 findings. Non-mention is mapped to negation in the reported template construction, which can introduce false-negative supervision. The reasoning and prominence fields are syn- thetic targets, not independent expert image annotations. This description summarizes the target structure and does not specify an exact reproducible prompt. S1.2 LLM-Based Label Expansion Pipeline Parsing and construction of the presence matrix. Each LLM response was parsed from its hierarchical JSON rep- resentation into tuples of (parent group, child tag, presence). Although the extraction schema additionally permitted attributes such as size, location, and characteristics, only the binary presence field was used for label construction and concordance analysis. Child labels were indexed jointly with their parent group to avoid collisions between identically named labels occurring in different sections of the schema. For a successfully parsed report, a tag explicitly marked as present was assigned a value of 1. Tags belonging to the model vocabulary but not emitted for that report were treated as implicitly absent (0). In contrast, reports for which the JSON response could not be parsed were treated as missing rather than negative and were excluded from the corresponding comparisons. This distinction prevented parsing failures from being interpreted as evidence for absence. A zero denotes the extraction convention and does not independently establish image-level absence. Vocabulary normalization. The predefined schema was supplemented with an open-ended other_findings field to preserve findings not anticipated by the original taxonomy. Consequently, the raw outputs did not share a perfectly fixed vocabulary. In the full concordance analysis, the seven models collectively produced 5,586 distinct keys, whereas only 193 canonical tags were emitted by all seven models; most of the remaining variation originated from other_findings. This expansion primarily reflected lexical fragmentation–for example, singular/plural forms such as chest_tube and chest_tubes–rather than thousands of distinct radiographic concepts. We therefore normalized the extracted vocabulary before constructing the final label space. Formatting variants and alternative representations of the same concept were first standardized, after which semantically equivalent labels were mapped to a canonical label. Importantly, normalization was performed at the level of clinical concepts rather than by string similarity alone. Labels were therefore merged only when they represented the same finding or a difference in naming granularity, while clinically related but distinct findings remained separate. Mapping to a parent group for concordance analysis should not be interpreted as evidence that the child concepts are interchangeable. 39 Synonym and hierarchical mapping. In the absence of ground-truth annotations for the expanded label set, concordance across multiple LLMs was used as a check of inter-model reproducibility. Agreement among models was interpreted as supporting evidence that the extracted labels were reproducible rather than as a mechanism for defining the la- bel mappings themselves. The concordance analysis also provided insight into the nature of residual disagreement. In particular, many apparent discrepancies occurred between synonymous or hierarchically related labels. For ex- ample, when solitary_pulmonary_nodule was not emitted, nodules was emitted in 96% of such cases; similarly, pleural_effusion was emitted in 95% of cases in which effusion was absent. These observations indicate that a substantial fraction of tag-level disagreement reflected differences in terminology or label granularity rather than failure to identify the underlying radiographic finding, and informed inspection of the normalization and mapping procedure without establishing clinical correctness. Fig. S1: Substitution patterns among discordant labels. For each missed tag (rows), cells show the probability that the same model assigned an alternative tag (columns) in the same report. High substitution rates identify cases where apparent disagreement reflects alternative label routing rather than failure to detect the underlying finding, as observed for effusion/pleural_effusion and the nodule family. We consequently examined related label families for normalization and parent-level concordance analysis. These families include both synonyms and related but non-equivalent findings; parent-level agreement does not establish equivalence at the child-label level. These included the nodule family; effusion/pleural_effusion; low-lung-volume variants; COPD-related terminology such as hyperinflation and emphysema; ILD/fibrotic terminology; and closely related edema terminology. The mapping was deliberately conservative. In particular, nodule, mass, and opacity were maintained as separate concepts despite their potential co-occurrence, since substitutions among these categories can represent genuine differences in radiographic interpretation rather than lexical variation. Similarly, the airspace- opacity family (e.g., consolidation, pneumonia, perihilar opacity, and edema) was not indiscriminately collapsed, as the concordance analysis identified this group as one of the remaining sources of substantive clinical disagreement. The effect of synonym routing was also assessed by comparing agreement at different levels of the hierarchy. Collaps- ing child labels into their parent finding groups increased mean pairwise Cohen’s κ from 0.83 to 0.87, indicating that a substantial component of child-level disagreement disappeared when differences in terminology and granularity were removed. Thus, parent-level agreement was used as an additional diagnostic for distinguishing genuine disagreement from alternative routing within the taxonomy. Handling free-form and ambiguous labels. Free-form other_findings labels received additional scrutiny because their vocabulary was substantially less standardized. Device and incidental findings were particularly affected: labels such as central venous catheters, PICC lines, chest tubes, surgical clips, and degenerative changes exhibited low apparent tag-level agreement despite frequently representing differences in naming or schema coverage. Free-form keys were 40 Fig. S2: Within-group correlation of extracted labels across the seven LLMs. Pairwise φ correlations between binary tag-presence indicators were computed across pooled model–report observations within each parent group. Higher φ indicates greater co-occurrence, capturing related, overlapping, or potentially synonymous labels. Fig. S3: Inter-model agreement and extraction characteristics across seven LLMs. Left: Pairwise Cohen’s κ at the child-label level, demonstrating consistently high agreement between models. Middle: Precision–recall char- acteristics relative to the seven-model majority consensus, illustrating differences in model extraction behavior from conservative (higher precision, lower recall) to more inclusive extraction. Right: Mean number of positive tags ex- tracted per report for each model; the dashed line indicates the seven-model mean. 41 therefore not automatically promoted to independent classes. Instead, recurrent keys were mapped to an existing canonical concept where possible; lexical variants were consolidated; and poorly supported or non-reproducible keys were flagged for taxonomy review. The artifacts label was handled separately because its disagreement pattern differed from synonym routing. Only 92 reports reached majority consensus for artifacts, although at least one model emitted the label in 228 reports. Moreover, models emitting artifacts generally also emitted the corresponding radiographic abnormalities, indicating that artifacts behaved as an additional, model-dependent catch-all rather than as an alternative name for another finding. Such labels were therefore not merged with co-occurring clinical findings and were instead treated as candidates for stricter definition or removal. Fig. S4: Model-specific behavior of the artifacts label. The artifacts label shows substantial model-specific variation and is particularly frequent for GLM 5.1. This reflects a model-specific labeling tendency in which GLM 5.1 often assigns the generic artifacts label alongside more specific findings, such as devices or tubes. The high frequency of single-model calls consequently suggests that artifacts represents an inconsistently interpreted label rather than systematic disagreement in the underlying radiographic findings. Taxonomy auditing and filtering. Following normalization, tags were audited using several complementary indicators of instability. These included the number of models containing the tag in their vocabulary, prevalence across models, pairwise agreement and Cohen’s κ, frequency of single-model positive calls, and the extent to which disagreement disappeared at the parent-group level. Tags occurring in only a minority of model vocabularies were interpreted primarily as schema-coverage differences rather than reliable independent categories. Finally, majority voting across the seven models was used during concordance analysis to identify systematic routing and extraction differences. A report-tag pair was considered consensus-positive when at least four of seven models marked it as present. This consensus served as an analytical reference rather than ground truth: over-calls and under- calls were defined relative to the ≥ 4/7 majority and were subsequently examined for systematic substitutions. This procedure was particularly useful for separating consensus-relative omissions from vocabulary routing; for example, the apparently frequent under-calling of effusion was largely attributable to models using pleural_effusion instead. Overall, the post-processing procedure was designed to reduce vocabulary fragmentation without artificially increasing agreement by collapsing clinically distinct findings. 42 EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BiomedVIL MedCLIP AR-2B (ours) AR-8B (ours) 0 20 40 60 80 100 Top-2 Win Rate (%) 0.9 0.0 27.0 20.0 0.9 1.7 0.0 45.2 14.8 89.6 0.00.0 42.6 16.7 0.00.00.0 8.3 40.7 91.7 0.00.0 37.9 24.1 0.00.00.0 6.9 93.1 37.9 AUROC Top-2 Win Rate CheXpert MIMIC-CXR PadChest Fig. S5: Top-2 Win Rate (%) under AUROC across CheXpert, MIMIC-CXR, and PadChest. Our Med-AR-2B and Med-AR-8B models achieve top-2 placements on a larger fraction of tags than prior encoders. Numeric values are annotated above each bar for clarity. S2 Cross-Dataset Summary and Operating Points S2.1 Operating-Point Metrics For public-dataset sensitivity, specificity, and Youden’s J, thresholds are selected per label by maximizing J = Sensitivity +Specificity−1 on the validation set. These operating points summarize the sensitivity–specificity trade-off, rather than a clinically validated deployment threshold. The internal per-label table is described in the source analysis as validation-set performance with class-wise J-maximizing thresholds; it should be distinguished from the held-out aggregate comparison in the main text. S2.2 Top-2 Win Rate Analysis for Public Datasets To provide a holistic view of model competitiveness across the entire label space of chest radiograph benchmarks, we summarize the per-tag performance of every method using a Top-2 Win Rate metric. For each tag in a given dataset, we rank all ten competing encoders by their score and award a “win” to any model that appears among the top two. The win rate of a model is then defined as the fraction of tags (expressed as a percentage) on which it achieves a top-2 placement. This formulation rewards models that are consistently strong across the full tag distribution, rather than those that excel on only a handful of well-represented findings. Figure S5 reports the Top-2 Win Rate measured under AUROC, while Figure S6 reports the same analysis under AUPRC. We evaluate on three widely used chest X-ray benchmarks CheXpert, MIMIC-CXR, and PadChest covering a broad spectrum of pathology labels ranging from common findings to long-tailed diagnoses. Med-AR-8B has the highest reported top-2 rates on CheXpert (89.6% AUROC, 85.2% AUPRC) and MIMIC-CXR (91.7% AUROC, 78.7% AUPRC), while Med-AR-2B leads on PadChest (93.1% AUROC, 96.6% AUPRC). This summary describes relative placement across labels; it does not measure effect size or establish rare-label clinical performance. Per-label values are provided by dataset in Sections S4–S6. 43 EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BiomedVIL MedCLIP AR-2B (ours) AR-8B (ours) 0 20 40 60 80 100 Top-2 Win Rate (%) 1.7 0.0 26.1 31.3 0.90.9 0.0 34.8 20.0 85.2 0.9 0.0 28.7 17.6 2.8 0.00.0 18.5 52.8 78.7 0.00.0 27.6 20.7 0.00.00.0 13.8 96.6 41.4 AuPRC Top-2 Win Rate CheXpert MIMIC-CXR PadChest Fig. S6: Top-2 Win Rate (%) under AUPRC across CheXpert, MIMIC-CXR, and PadChest. Consistent with the AUROC results, our proposed models lead on precision-recall performance, with Med-AR-2B particularly excelling on the long-tailed PadChest benchmark. 44 S3 Internal Dataset: Per-Label ResultsThe following table preserves the reported validation-set results for 40 labels. AUROC and AUPRC describe discrimination; sensitivity, specificity, andYouden’s J describe class-wise operating points. Performance varies across labels and metrics, with neither pretraining recipe uniformly superior. Table S1: Consolidated per-label performance on the internal dataset across AUROC, AUPRC, Sensitivity, Specificity, and Youden’s J for Med-CLIP and our autoregressive models (AR-2B, AR-8B). Thresholds for Sensitivity/Specificity/Youden are obtained per abnormality by maximising J = Sens + Spec − 1 . Tail abnormalities are marked with an asterisk (*), and medium-prevalence abnormalities with a dagger ( † ). For each metric, the best value per row is shown in bold and the second-best is underlined. Abnormality AUROC AUPRC Sensitivity Specificity Youden’s J Med-CLIP AR-2B (ours) AR-8B (ours) Med-CLIP AR-2B (ours) AR-8B (ours) Med-CLIP AR-2B (ours) AR-8B (ours) Med-CLIP AR-2B (ours) AR-8B (ours) Med-CLIP AR-2B (ours) AR-8B (ours) Hydropneumothorax ∗ 0.9899 0.9888 0.9863 0.4445 0.3672 0.3013 0.9641 0.9596 0.9417 0.9634 0.9623 0.9713 0.9275 0.9219 0.9130 Hilar Mass ∗ 0.8907 0.8904 0.8764 0.3714 0.3275 0.3671 0.9107 0.9107 0.8571 0.7424 0.7285 0.7850 0.6531 0.6392 0.6421 Granuloma † 0.9185 0.9481 0.9367 0.0782 0.0863 0.0804 0.8346 0.8508 0.8612 0.8395 0.9091 0.8815 0.6741 0.7599 0.7427 Consolidation 0.8687 0.8902 0.8877 0.5193 0.5746 0.5698 0.7924 0.8128 0.8142 0.7824 0.8091 0.8026 0.5748 0.6219 0.6168 Lesion † 0.5796 0.5431 0.5764 0.0792 0.0744 0.0810 0.6790 0.7433 0.6639 0.4683 0.3474 0.4596 0.1473 0.0907 0.1235 Fibrosis 0.8588 0.8866 0.8821 0.3234 0.3986 0.3849 0.7578 0.7951 0.7865 0.7937 0.8168 0.8131 0.5515 0.6119 0.5996 Scoliosis † 0.8772 0.9162 0.9107 0.2814 0.3580 0.3536 0.7769 0.8056 0.7859 0.8583 0.8867 0.9065 0.6352 0.6923 0.6924 Pleural Plaque ∗ 0.9516 0.9639 0.9475 0.1115 0.0767 0.0679 0.9145 0.9145 0.8547 0.8719 0.9326 0.9475 0.7864 0.8471 0.8022 Clavicle Fracture † 0.9211 0.9065 0.8811 0.4177 0.3338 0.2714 0.8033 0.7912 0.7495 0.9381 0.8933 0.8787 0.7414 0.6845 0.6282 Opacity 0.8576 0.8652 0.8618 0.6665 0.6872 0.6899 0.7772 0.7495 0.7762 0.7759 0.8218 0.7920 0.5531 0.5713 0.5682 Cardiomegaly 0.9336 0.9319 0.9322 0.4962 0.4758 0.4906 0.8997 0.8800 0.8916 0.8219 0.8414 0.8314 0.7216 0.7214 0.7230 Rib Fracture † 0.9215 0.9252 0.9141 0.2046 0.2226 0.1909 0.8190 0.8253 0.8121 0.8672 0.8708 0.8561 0.6862 0.6961 0.6682 Pneumothorax † 0.9708 0.9736 0.9722 0.5327 0.5774 0.5313 0.8738 0.8868 0.8991 0.9503 0.9586 0.9377 0.8241 0.8454 0.8368 Mediastinal Mass ∗ 0.9724 0.9891 0.9750 0.2318 0.1814 0.2082 0.8800 0.9400 0.9467 0.9811 0.9544 0.9554 0.8611 0.8944 0.9021 Cyst ∗ 0.8566 0.8868 0.9137 0.0468 0.0494 0.0550 0.7976 0.7614 0.8265 0.7716 0.8625 0.8607 0.5692 0.6239 0.6872 Ground Glass ∗ 0.8830 0.8683 0.8614 0.0261 0.0165 0.0227 0.8051 0.8159 0.7906 0.8160 0.7800 0.7558 0.6211 0.5959 0.5464 Reticulonodular Pattern † 0.8417 0.8711 0.8588 0.1314 0.1334 0.1246 0.6960 0.7469 0.7654 0.8363 0.8283 0.8033 0.5323 0.5752 0.5687 Pneumonia 0.7784 0.7991 0.7944 0.1157 0.1309 0.1254 0.6299 0.6847 0.7104 0.8039 0.7587 0.7199 0.4338 0.4434 0.4303 Nodule 0.8910 0.9110 0.9066 0.3490 0.4264 0.4062 0.8143 0.8275 0.8112 0.8154 0.8493 0.8540 0.6297 0.6768 0.6652 Scarring ∗ 0.8923 0.9211 0.9271 0.0587 0.0644 0.0759 0.7771 0.8471 0.8726 0.8643 0.8562 0.8598 0.6414 0.7033 0.7324 Pleural Effusion 0.9369 0.9507 0.9485 0.5978 0.6813 0.6728 0.8838 0.9034 0.8916 0.8477 0.8626 0.8682 0.7315 0.7660 0.7598 Mass ∗ 0.9879 0.9922 0.9852 0.2868 0.2856 0.2849 0.9532 0.9640 0.9353 0.9463 0.9531 0.9600 0.8995 0.9171 0.8953 Mediastinal Shift † 0.9447 0.9667 0.9563 0.2132 0.2128 0.2166 0.8424 0.8923 0.9014 0.9193 0.9252 0.8878 0.7617 0.8175 0.7892 Pulmonary Edema † 0.8918 0.9070 0.8946 0.0919 0.0601 0.0681 0.7873 0.8171 0.8260 0.8336 0.8343 0.8086 0.6209 0.6514 0.6346 Hilar Prominence 0.7663 0.7627 0.7535 0.1057 0.0946 0.0940 0.6712 0.6798 0.6631 0.7168 0.6802 0.6966 0.3880 0.3600 0.3597 Bronchitis 0.7573 0.7724 0.7653 0.2653 0.2698 0.2657 0.6225 0.7048 0.6613 0.7561 0.6876 0.7264 0.3786 0.3924 0.3877 Emphysema † 0.9025 0.9136 0.9065 0.2231 0.2101 0.2216 0.7656 0.8243 0.8229 0.8778 0.8531 0.8360 0.6434 0.6774 0.6589 Calcification 0.9188 0.9337 0.9288 0.3730 0.4181 0.4097 0.8274 0.8464 0.8445 0.8695 0.8780 0.8712 0.6969 0.7244 0.7157 Cavity † 0.9459 0.9579 0.9512 0.1694 0.3671 0.2487 0.8715 0.8772 0.8723 0.9055 0.9224 0.9189 0.7770 0.7996 0.7912 Lymphadenopathy ∗ 0.6685 0.6592 0.5931 0.0128 0.0121 0.0098 0.6667 0.7187 0.6496 0.5854 0.5388 0.4933 0.2521 0.2575 0.1429 Hernia † 0.9672 0.9686 0.9672 0.1406 0.1373 0.1426 0.9067 0.9300 0.9026 0.9233 0.9191 0.9447 0.8300 0.8491 0.8473 Normal 0.8549 0.8548 0.8562 0.8542 0.8524 0.8589 0.7367 0.7442 0.7237 0.7863 0.7716 0.7959 0.5230 0.5158 0.5196 Pleural Thickening 0.9090 0.9128 0.9108 0.1522 0.1751 0.1563 0.8504 0.8201 0.8279 0.8192 0.8528 0.8361 0.6696 0.6729 0.6640 Edema ∗ 0.9338 0.9518 0.9457 0.1266 0.0855 0.0847 0.8669 0.8912 0.8523 0.8763 0.8892 0.9081 0.7432 0.7804 0.7604 Tented Diaphragm ∗ 0.8726 0.8962 0.9080 0.0295 0.0181 0.0275 0.6820 0.7972 0.8479 0.9231 0.8387 0.8272 0.6051 0.6359 0.6751 Osteopenia † 0.9223 0.9352 0.9411 0.0993 0.0925 0.1134 0.8592 0.9224 0.8881 0.8363 0.7910 0.8408 0.6955 0.7134 0.7289 Atelectasis 0.9426 0.9532 0.9491 0.1978 0.3358 0.3290 0.8902 0.9131 0.9018 0.8625 0.8632 0.8643 0.7527 0.7763 0.7661 Mediastinal Widening ∗ 0.9347 0.9390 0.9389 0.0831 0.0955 0.0977 0.8095 0.8701 0.8225 0.9097 0.8622 0.9098 0.7192 0.7323 0.7323 Metastasis ∗ 0.9528 0.9569 0.9205 0.5488 0.5417 0.4333 0.9242 0.9242 0.9091 0.8564 0.8763 0.7824 0.7806 0.8005 0.6915 Tracheal Shift † 0.9441 0.9416 0.9428 0.2555 0.2492 0.2389 0.8523 0.8376 0.8404 0.9213 0.9315 0.9362 0.7736 0.7691 0.7766 45 S4 CheXpert: Per-Label ResultsThe following tables report the existing results for 115 labels, in the order AUROC, AUPRC, sensitivity, specificity, and Youden’s J . S4.1 AUROCTable S2: Per-label AUROC comparison across baseline, contrastive, and autoregressive models on the CheXpert dataset. Tail abnormalities are marked withan asterisk (*), and medium-prevalence abnormalities are marked with a dagger ( † ). For each row, the best value is shown in bold and the second-best value is underlined. Abnormality EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BioViL-T Med-CLIP Med-AR-2B (ours) Med-AR-8B (ours) LVAD* 0.9166 0.9202 0.9499 0.9741 0.8624 0.9086 0.6904 0.9581 0.9579 0.9893 Diaphragm Elevation 0.7963 0.7187 0.8422 0.8996 0.7299 0.7362 0.5589 0.8998 0.8808 0.9134 Osteopenia † 0.9038 0.8549 0.8978 0.9006 0.8195 0.8803 0.6719 0.9102 0.9052 0.9130 Parenchymal Opacities † 0.7188 0.7463 0.7707 0.7331 0.6791 0.7871 0.7458 0.7794 0.7440 0.7681 Nodules 0.7665 0.7578 0.8370 0.8083 0.7394 0.7590 0.7210 0.8102 0.7904 0.8347 Pleural Effusion 0.8818 0.8527 0.9083 0.9089 0.8191 0.8817 0.7735 0.9053 0.8921 0.9075 Congestive Heart Failure* 0.7978 0.7765 0.8512 0.8580 0.7945 0.8033 0.5773 0.8646 0.8707 0.8667 Prosthetic Valve* 0.8379 0.8322 0.8643 0.8677 0.7147 0.8424 0.5762 0.8577 0.8327 0.8779 Support Devices 0.8492 0.8302 0.8552 0.8551 0.8184 0.8448 0.8160 0.8535 0.8493 0.8595 Picc Line 0.7111 0.6496 0.9529 0.9516 0.6409 0.6623 0.5440 0.9087 0.9531 0.9597 Pleural Thickening 0.7690 0.7497 0.8239 0.8286 0.7269 0.7693 0.6562 0.8395 0.8117 0.8541 Soft Tissue Swelling † 0.8643 0.7971 0.8858 0.9126 0.7717 0.8488 0.4615 0.8981 0.8813 0.9077 Neoplastic Nodules* 0.8542 0.8063 0.9550 0.9495 0.8143 0.8330 0.6700 0.9485 0.9000 0.9682 Nasogastric Tube Removal* 0.7674 0.7300 0.8037 0.8297 0.7558 0.7099 0.6312 0.8351 0.8107 0.8428 Venous Catheter* 0.7332 0.7269 0.7296 0.7899 0.6614 0.7296 0.6284 0.8028 0.8334 0.8364 Hemothorax* 0.7830 0.7981 0.8600 0.8110 0.6953 0.7512 0.6220 0.8817 0.8264 0.8593 CABG † 0.7911 0.7746 0.8238 0.8310 0.7076 0.7513 0.5187 0.8185 0.8382 0.8391 Right Subclavian Line* 0.7321 0.7359 0.7880 0.7820 0.6879 0.7312 0.6872 0.8387 0.8121 0.9119 Aortic Tortuosity † 0.8860 0.8632 0.9102 0.9229 0.8407 0.8722 0.7381 0.9273 0.9286 0.9424 Solitary Nodules † 0.7977 0.7841 0.8548 0.8276 0.7541 0.7829 0.7919 0.8400 0.8332 0.8679 Feeding Tube Removal* 0.6791 0.5639 0.6685 0.7191 0.6868 0.6166 0.6035 0.6845 0.6169 0.7466 Scoliosis † 0.8366 0.7468 0.8232 0.9253 0.7545 0.7590 0.6134 0.9033 0.8589 0.9415 Pulmonary Edema 0.8501 0.8240 0.8706 0.8682 0.7798 0.8457 0.7574 0.8656 0.8492 0.8708 Rib Fractures 0.7168 0.6575 0.8315 0.8213 0.6639 0.6666 0.5499 0.8265 0.7712 0.8307 Ards* 0.9200 0.9055 0.9273 0.9214 0.8517 0.9148 0.8628 0.9240 0.9180 0.9412 Aspiration 0.7089 0.6604 0.7656 0.7798 0.6318 0.6978 0.5871 0.7653 0.7594 0.7809 Reticular Opacities 0.7730 0.7351 0.8075 0.8053 0.6610 0.7793 0.5444 0.8091 0.7833 0.8150 Kyphosis* 0.8945 0.8348 0.9259 0.9522 0.8384 0.8347 0.7309 0.9409 0.9184 0.9500 Vascular Redistribution Cephalization † 0.8127 0.7749 0.8071 0.8011 0.7393 0.7733 0.5645 0.8197 0.7785 0.8434 Airspace Opacity † 0.7196 0.6994 0.7391 0.7412 0.6701 0.7446 0.6788 0.7319 0.6848 0.7347 Left Subclavian Catheter* 0.6514 0.5566 0.7941 0.8431 0.6635 0.5646 0.5941 0.8248 0.9000 0.9354 Kerley B Lines* 0.8169 0.8084 0.8699 0.8742 0.6992 0.7844 0.5527 0.8879 0.8641 0.8944 Chest Tube Removal † 0.7925 0.7513 0.8175 0.8695 0.7198 0.7345 0.5142 0.8546 0.8145 0.8822 Degenerative Changes † 0.8176 0.7939 0.8305 0.8434 0.7774 0.8068 0.7154 0.8477 0.8436 0.8518 Dialysis Catheter* 0.8084 0.7168 0.8797 0.8912 0.8335 0.6677 0.5829 0.9190 0.8968 0.9592 Hilar Lymphadenopathy* 0.7395 0.7471 0.8465 0.8085 0.6727 0.6885 0.6349 0.8773 0.7914 0.8916 Pigtail Catheter † 0.7543 0.7104 0.9481 0.8481 0.6950 0.7274 0.5755 0.8968 0.9251 0.9786 Epidural Catheter † 0.9258 0.8821 0.9510 0.9674 0.8879 0.8669 0.5601 0.9836 0.9829 0.9908 Mediastinal Drain † 0.9368 0.9215 0.9517 0.9542 0.8753 0.9188 0.7046 0.9482 0.9431 0.9644 Retrocardiac Opacity 0.7578 0.7042 0.7830 0.7843 0.6918 0.7380 0.6490 0.7893 0.7686 0.8019 Continued on next page 46 Abnormality EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BioViL-T Med-CLIP Med-AR-2B (ours) Med-AR-8B (ours) Pneumothorax 0.8477 0.8096 0.9410 0.9145 0.7518 0.8585 0.6546 0.9151 0.9103 0.9366 Prosthetic Aortic Valve* 0.7845 0.8111 0.8559 0.8565 0.7901 0.8092 0.6247 0.8469 0.8572 0.8820 Extubation † 0.8111 0.7902 0.8257 0.8268 0.7756 0.7599 0.7213 0.8608 0.8684 0.8797 Perihilar Opacities 0.7605 0.7402 0.8297 0.8354 0.6927 0.7870 0.6533 0.8424 0.8073 0.8463 Sternotomy Wires 0.9020 0.8885 0.9010 0.9044 0.8191 0.8938 0.5702 0.9050 0.8970 0.9105 Subcutaneous Emphysema † 0.9203 0.8292 0.9436 0.9552 0.7612 0.8992 0.5771 0.9532 0.9493 0.9568 MPA Enlargement † 0.7945 0.7179 0.8028 0.8886 0.6877 0.7504 0.5440 0.8939 0.8487 0.9023 Bronchial Wall Thickening* 0.8241 0.7592 0.8106 0.8211 0.7751 0.7826 0.7107 0.8349 0.7931 0.8118 Chest Wall Deformity † 0.6937 0.6386 0.7491 0.7828 0.6113 0.6535 0.6083 0.7968 0.7399 0.8193 Feeding Tube 0.9345 0.8939 0.9453 0.9474 0.8836 0.9358 0.8271 0.9501 0.9489 0.9540 Tracheostomy Cannula* 0.8992 0.7933 0.9747 0.9772 0.7970 0.8382 0.7369 0.9734 0.9736 0.9771 Sternotomy † 0.8674 0.8578 0.8704 0.8735 0.7868 0.8538 0.4712 0.8725 0.8718 0.8901 Granuloma † 0.7746 0.7485 0.8632 0.8020 0.7422 0.7537 0.7262 0.8884 0.8780 0.9194 Mediastinal Mass* 0.6562 0.7037 0.8533 0.8446 0.6021 0.7054 0.6528 0.8561 0.8506 0.8827 Basilar Opacity 0.7283 0.7103 0.7543 0.7613 0.6992 0.7273 0.6826 0.7560 0.7580 0.7699 Artifacts 0.6522 0.5949 0.6659 0.6870 0.6047 0.6206 0.5205 0.6928 0.6690 0.7007 Cardiomegaly 0.8725 0.8181 0.8913 0.8901 0.7813 0.8194 0.6058 0.8895 0.8752 0.8950 Solitary Pulmonary Nodule † 0.7711 0.7583 0.7943 0.7762 0.7377 0.7750 0.7496 0.7996 0.7978 0.8236 Mediastinal Widening 0.8001 0.6869 0.8382 0.8498 0.7043 0.7069 0.5373 0.8620 0.8368 0.8759 Improved Aeration* 0.5931 0.5553 0.5884 0.6413 0.6301 0.5681 0.5315 0.5652 0.5860 0.6239 Chest Tube 0.9264 0.8918 0.9532 0.9467 0.8278 0.8920 0.6535 0.9437 0.9479 0.9599 Malignant Nodules* 0.8661 0.8115 0.9555 0.9452 0.8211 0.8187 0.6661 0.9324 0.8826 0.9589 Tracheostomy Tube † 0.9555 0.8370 0.9756 0.9759 0.8993 0.9425 0.7401 0.9759 0.9749 0.9801 Cancer 0.7715 0.7473 0.8309 0.8358 0.7147 0.7495 0.6059 0.8370 0.8276 0.8650 Pacemaker † 0.9732 0.9419 0.9726 0.9738 0.9295 0.9368 0.5153 0.9755 0.9748 0.9827 Mediastinal Drains † 0.9493 0.9387 0.9620 0.9570 0.9003 0.9330 0.7331 0.9595 0.9577 0.9674 Atrial Dilation* 0.7600 0.7669 0.7952 0.7698 0.7388 0.5822 0.6098 0.7696 0.7613 0.7988 AICD † 0.9665 0.9618 0.9723 0.9789 0.9545 0.9565 0.6053 0.9801 0.9780 0.9818 Subclavian Line* 0.6966 0.6659 0.7601 0.8333 0.6128 0.6588 0.5865 0.8324 0.8795 0.9429 Emphysema † 0.9094 0.8341 0.9233 0.9300 0.8131 0.8945 0.5888 0.9298 0.9253 0.9446 Interstitial Edema 0.7569 0.7259 0.7921 0.7759 0.6852 0.7563 0.6126 0.7870 0.7546 0.7933 Hydropneumothorax* 0.9014 0.8390 0.9338 0.9077 0.7321 0.8786 0.6357 0.9511 0.9162 0.9530 Diaphragmatic Hernia † 0.8113 0.7611 0.8312 0.8982 0.7567 0.7702 0.5783 0.8661 0.8632 0.9059 Pericardial Effusion* 0.7655 0.7115 0.8452 0.8197 0.7153 0.6875 0.4885 0.8543 0.8313 0.8420 Healed Rib Fracture † 0.7463 0.7193 0.8675 0.8342 0.7207 0.7326 0.6600 0.8810 0.7669 0.8792 Alveolar Edema † 0.8627 0.8465 0.8941 0.8854 0.7921 0.8668 0.7051 0.8908 0.8645 0.8913 Aortic Valve Replacement* 0.8268 0.8159 0.8767 0.8775 0.7777 0.8216 0.6329 0.8692 0.8545 0.8915 Hilar Enlargement † 0.6956 0.6542 0.8153 0.8710 0.6114 0.6671 0.5968 0.8591 0.8317 0.8830 Internal Jugular Line 0.8386 0.7987 0.8828 0.8898 0.7750 0.8598 0.6799 0.8904 0.8837 0.8963 Normal 0.9114 0.9061 0.9225 0.9225 0.8746 0.9059 0.8818 0.9211 0.9172 0.9205 Diffused Nodules † 0.8409 0.8076 0.9506 0.9328 0.8166 0.8269 0.6626 0.9393 0.9256 0.9492 Calcified Nodules* 0.7685 0.7498 0.8400 0.7680 0.6903 0.7797 0.7527 0.7950 0.7914 0.8796 Clavicle Fracture † 0.7520 0.7261 0.7847 0.7645 0.6684 0.6941 0.6106 0.9335 0.7440 0.8008 Nasogastric Tube 0.9335 0.9263 0.9464 0.9480 0.9013 0.9413 0.9001 0.9488 0.9432 0.9560 Central Venous Line 0.7840 0.7451 0.8450 0.8437 0.7236 0.7934 0.5626 0.8373 0.8503 0.8599 Left Subclavian Line † 0.7765 0.7325 0.8341 0.8974 0.7124 0.7367 0.6033 0.8648 0.9304 0.9511 Pneumomediastinum* 0.8005 0.6861 0.8727 0.8218 0.7725 0.7776 0.3785 0.8128 0.8096 0.8873 Interstitial Thickening † 0.7520 0.7397 0.7763 0.7583 0.6825 0.7668 0.5192 0.7738 0.7239 0.7719 Pulmonary Vascular Congestion 0.7103 0.6779 0.7402 0.7268 0.6502 0.7001 0.5697 0.7309 0.6984 0.7392 Swan Ganz Catheter 0.8835 0.8768 0.9522 0.9503 0.9183 0.9163 0.7876 0.9552 0.9516 0.9657 Aeration* 0.6051 0.6036 0.5738 0.5811 0.5481 0.5210 0.4089 0.5802 0.6306 0.6224 Mediastinal Shift* 0.9134 0.8151 0.9191 0.9515 0.8745 0.8350 0.5067 0.9576 0.9278 0.9572 Surgical Hardware* 0.6378 0.6314 0.7024 0.6091 0.5865 0.6159 0.4780 0.7109 0.7502 0.7558 Patchy Opacities* 0.7612 0.7208 0.7640 0.7655 0.6777 0.7832 0.6709 0.8215 0.8000 0.8194 Continued on next page 47 Abnormality EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BioViL-T Med-CLIP Med-AR-2B (ours) Med-AR-8B (ours) Mediastinal Clips † 0.8610 0.8498 0.8852 0.8988 0.7953 0.8347 0.4249 0.9009 0.8895 0.9118 Opacity 0.6974 0.6762 0.7278 0.7218 0.6487 0.6814 0.6495 0.7243 0.7067 0.7261 Pulmonary Hypertension* 0.7483 0.6552 0.7356 0.8282 0.6695 0.6649 0.5263 0.8486 0.7865 0.8534 Atelectasis 0.6999 0.6854 0.7406 0.7341 0.6449 0.7019 0.6232 0.7304 0.7111 0.7361 Hyperinflation † 0.7043 0.6969 0.7057 0.7292 0.6627 0.6827 0.6253 0.7342 0.6750 0.7362 ILD 0.7413 0.7030 0.7797 0.7758 0.6572 0.7386 0.5627 0.7777 0.7481 0.7869 COPD † 0.7727 0.7650 0.8151 0.8076 0.7281 0.8003 0.6325 0.8196 0.8016 0.8169 Opacification* 0.8351 0.7904 0.8739 0.8135 0.8401 0.7490 0.6582 0.8506 0.7979 0.8676 Mediport* 0.8160 0.7758 0.9429 0.9307 0.7324 0.6849 0.5639 0.9564 0.9703 0.9884 Lung Transplant* 0.8811 0.8037 0.9180 0.9320 0.7957 0.8303 0.5093 0.9282 0.9389 0.9472 Surgical Clips 0.6955 0.6398 0.8021 0.7760 0.6624 0.6233 0.5709 0.7975 0.8346 0.8892 Low Lung Volumes 0.8326 0.8131 0.8411 0.8428 0.7836 0.8111 0.6937 0.8443 0.8283 0.8436 Aortic Aneurysm † 0.7383 0.6037 0.7917 0.8449 0.6906 0.6250 0.5755 0.8727 0.8700 0.9086 Consolidation 0.7348 0.7214 0.7629 0.7558 0.6981 0.7347 0.6911 0.7517 0.7344 0.7579 Pneumonia 0.7344 0.6828 0.7883 0.7855 0.6433 0.7293 0.5692 0.7727 0.7639 0.7853 Free Subdiaphragmatic Air* 0.7206 0.6754 0.7390 0.8138 0.6698 0.6819 0.6444 0.9159 0.8371 0.9286 Fibrosis 0.8414 0.8053 0.8896 0.8893 0.7710 0.8290 0.7200 0.8911 0.8751 0.9016 Lung Mass † 0.8102 0.7449 0.8947 0.8906 0.7255 0.7549 0.6157 0.9007 0.8739 0.9098 Endotracheal Tube 0.9454 0.9282 0.9503 0.9541 0.9259 0.9431 0.9163 0.9542 0.9539 0.9598 Pulmonary Arterial Hypertension Signs* 0.8105 0.7498 0.7922 0.8771 0.7009 0.7058 0.6119 0.8895 0.8120 0.8539 Calcification 0.8107 0.7701 0.8439 0.8555 0.7424 0.7796 0.6371 0.8657 0.8341 0.8872 48 S4.2 AUPRCTable S3: Per-label AUPRC comparison across baseline, contrastive, and autoregressive models on the CheXpert dataset. Tail abnormalities are marked withan asterisk (*), and medium-prevalence abnormalities are marked with a dagger ( † ). For each row, the best value is shown in bold and the second-best value is underlined. Abnormality EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BioViL-T Med-CLIP Med-AR-2B (ours) Med-AR-8B (ours) LVAD* 0.0814 0.0851 0.1213 0.2439 0.0848 0.0591 0.0093 0.1771 0.2239 0.3616 Diaphragm Elevation 0.1911 0.1063 0.2607 0.4024 0.1548 0.1251 0.0426 0.4165 0.4013 0.4314 Osteopenia † 0.1545 0.0869 0.1384 0.1384 0.0884 0.1144 0.0248 0.1508 0.1462 0.1724 Parenchymal Opacities † 0.0386 0.0139 0.0139 0.0206 0.0126 0.0255 0.0229 0.0217 0.0168 0.0209 Nodules 0.1140 0.1039 0.2721 0.2329 0.0906 0.1124 0.0725 0.2306 0.2164 0.2727 Pleural Effusion 0.8558 0.8127 0.8902 0.8892 0.7774 0.8552 0.7057 0.8850 0.8715 0.8891 Congestive Heart Failure* 0.0218 0.0234 0.0236 0.0355 0.0172 0.0180 0.0058 0.0602 0.0362 0.0431 Prosthetic Valve* 0.0159 0.0152 0.0172 0.0188 0.0122 0.0111 0.0044 0.0169 0.0182 0.0155 Support Devices 0.2944 0.2592 0.2946 0.2947 0.2518 0.2889 0.2506 0.2880 0.2936 0.3110 Picc Line 0.1639 0.1377 0.6090 0.5874 0.1464 0.1436 0.0897 0.5075 0.5904 0.6169 Pleural Thickening 0.1147 0.0952 0.1542 0.1747 0.0889 0.0985 0.0456 0.1615 0.1578 0.1695 Soft Tissue Swelling † 0.3035 0.1142 0.3336 0.3315 0.1527 0.2508 0.0155 0.3401 0.2831 0.3211 Neoplastic Nodules* 0.0958 0.0403 0.2792 0.2552 0.0548 0.0301 0.0067 0.2977 0.1945 0.3072 Nasogastric Tube Removal* 0.0271 0.0203 0.0357 0.0304 0.0170 0.0151 0.0071 0.0423 0.0272 0.0328 Venous Catheter* 0.0078 0.0096 0.0121 0.0283 0.0073 0.0081 0.0066 0.0194 0.0194 0.0204 Hemothorax* 0.0237 0.0165 0.0705 0.0682 0.0265 0.0096 0.0052 0.0720 0.1020 0.1018 CABG † 0.0691 0.0610 0.0989 0.1155 0.0625 0.0563 0.0253 0.1038 0.1059 0.1214 Right Subclavian Line* 0.0088 0.0099 0.0122 0.0131 0.0071 0.0101 0.0073 0.0199 0.0536 0.0685 Aortic Tortuosity † 0.0996 0.0687 0.1290 0.1623 0.0634 0.0750 0.0290 0.1696 0.1580 0.2187 Solitary Nodules † 0.0195 0.0174 0.0415 0.0495 0.0208 0.0186 0.0166 0.0417 0.0513 0.0701 Feeding Tube Removal* 0.0077 0.0049 0.0076 0.0090 0.0069 0.0053 0.0048 0.0072 0.0050 0.0078 Scoliosis † 0.1196 0.0464 0.1283 0.3646 0.0689 0.0388 0.0178 0.3252 0.2158 0.4447 Pulmonary Edema 0.7112 0.6635 0.7504 0.7461 0.6067 0.7010 0.5489 0.7427 0.7158 0.7505 Rib Fractures 0.1042 0.0619 0.2604 0.2347 0.0791 0.0631 0.0402 0.2865 0.1842 0.2578 ARDS* 0.0742 0.0468 0.0724 0.0478 0.0277 0.0498 0.0154 0.0816 0.0681 0.1012 Aspiration 0.0699 0.0419 0.0839 0.0962 0.0380 0.0522 0.0293 0.0928 0.0773 0.1059 Reticular Opacities 0.2354 0.1840 0.2980 0.2996 0.1617 0.2506 0.0809 0.3062 0.2677 0.3118 Kyphosis* 0.0755 0.0405 0.0598 0.1064 0.0314 0.0255 0.0107 0.1672 0.1158 0.1658 Vascular Redistribution Cephalization † 0.0425 0.0281 0.0432 0.0515 0.0249 0.0290 0.0082 0.0457 0.0373 0.0435 Airspace Opacity † 0.0303 0.0309 0.0490 0.0499 0.0244 0.0378 0.0259 0.0415 0.0235 0.0491 Left Subclavian Catheter* 0.0101 0.0045 0.0141 0.0391 0.0079 0.0056 0.0043 0.0256 0.0352 0.0412 Kerley B Lines* 0.0367 0.0186 0.0454 0.0723 0.0108 0.0195 0.0034 0.1033 0.0287 0.0666 Chest Tube Removal † 0.0597 0.0375 0.0612 0.0866 0.0244 0.0231 0.0060 0.0831 0.0523 0.1153 Degenerative Changes † 0.0953 0.0836 0.1081 0.1205 0.0843 0.0945 0.0478 0.1194 0.1188 0.1262 Dialysis Catheter* 0.0240 0.0067 0.0408 0.0899 0.0284 0.0061 0.0043 0.0834 0.1176 0.1352 Hilar Lymphadenopathy* 0.0119 0.0143 0.0678 0.1143 0.0087 0.0113 0.0074 0.0617 0.0506 0.1172 Pigtail Catheter † 0.0192 0.0174 0.1676 0.0988 0.0211 0.0143 0.0123 0.1614 0.1593 0.1983 Epidural Catheter † 0.1511 0.1151 0.2646 0.2664 0.1618 0.0623 0.0108 0.3916 0.4249 0.4096 Mediastinal Drain † 0.2022 0.1868 0.2301 0.2504 0.1058 0.2036 0.0274 0.2231 0.2130 0.3026 Retrocardiac Opacity 0.1270 0.0982 0.1464 0.1501 0.0905 0.1121 0.0681 0.1589 0.1406 0.1684 Pneumothorax 0.4286 0.3569 0.7242 0.5663 0.3012 0.4442 0.1329 0.6153 0.6045 0.6964 Prosthetic Aortic Valve* 0.0114 0.0116 0.0208 0.0180 0.0144 0.0124 0.0074 0.0417 0.0307 0.0315 Extubation † 0.0640 0.0458 0.0680 0.0646 0.0362 0.0324 0.0237 0.0761 0.0743 0.0885 Perihilar Opacities 0.0866 0.0755 0.1626 0.1560 0.0583 0.1010 0.0415 0.1617 0.1617 0.1857 Sternotomy Wires 0.3066 0.2770 0.2993 0.3168 0.2202 0.2883 0.0810 0.3119 0.2957 0.3266 Subcutaneous Emphysema † 0.1897 0.0705 0.2326 0.2729 0.0856 0.1534 0.0166 0.2661 0.2471 0.2750 Continued on next page 49 Abnormality EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BioViL-T Med-CLIP Med-AR-2B (ours) Med-AR-8B (ours) MPA Enlargement † 0.0431 0.0185 0.0678 0.2743 0.0330 0.0281 0.0087 0.2455 0.2870 0.3013 Bronchial Wall Thickening* 0.0344 0.0280 0.0621 0.0451 0.0213 0.0262 0.0133 0.0533 0.0360 0.0489 Chest Wall Deformity † 0.0374 0.0200 0.0431 0.0468 0.0163 0.0218 0.0163 0.0450 0.0653 0.0621 Feeding Tube 0.4000 0.2994 0.4684 0.4868 0.3185 0.4258 0.1932 0.4900 0.4891 0.5157 Tracheostomy Cannula* 0.0386 0.0100 0.1304 0.1068 0.0659 0.0328 0.0074 0.0743 0.1089 0.1111 Sternotomy † 0.0504 0.0444 0.0523 0.0534 0.0445 0.0448 0.0117 0.0492 0.0509 0.0660 Granuloma † 0.0513 0.0397 0.1542 0.0771 0.0402 0.0460 0.0299 0.2658 0.2784 0.3594 Mediastinal Mass* 0.0107 0.0109 0.0640 0.1148 0.0277 0.0096 0.0098 0.0928 0.1136 0.1231 Basilar Opacity 0.1363 0.1265 0.1547 0.1592 0.1300 0.1388 0.1103 0.1578 0.1663 0.1677 Artifacts 0.1236 0.0859 0.1328 0.1652 0.0952 0.1019 0.0668 0.1539 0.1588 0.1734 Cardiomegaly 0.6509 0.5194 0.6820 0.6802 0.4937 0.5205 0.2464 0.6783 0.6560 0.6889 Solitary Pulmonary Nodule † 0.0232 0.0256 0.0540 0.0414 0.0217 0.0293 0.0223 0.0479 0.0511 0.0761 Mediastinal Widening 0.1519 0.0550 0.2140 0.2497 0.0734 0.0655 0.0327 0.2669 0.2529 0.2916 Improved Aeration* 0.0097 0.0088 0.0092 0.0124 0.0124 0.0086 0.0086 0.0082 0.0089 0.0143 Chest Tube 0.5163 0.4454 0.5913 0.5621 0.3419 0.3882 0.1051 0.5720 0.5898 0.6285 Malignant Nodules* 0.0836 0.0254 0.2941 0.2682 0.0524 0.0281 0.0064 0.2978 0.2060 0.3151 Tracheostomy Tube † 0.3432 0.1055 0.4711 0.4581 0.2830 0.3519 0.0591 0.4678 0.4768 0.4943 Cancer 0.1184 0.0855 0.2157 0.2118 0.0756 0.0914 0.0372 0.2199 0.2064 0.2671 Pacemaker † 0.3781 0.2249 0.4018 0.4577 0.3505 0.2188 0.0211 0.4368 0.4423 0.4977 Mediastinal Drains † 0.1196 0.0793 0.1648 0.1329 0.0771 0.0692 0.0183 0.1385 0.1318 0.1639 Atrial Dilation* 0.0190 0.0178 0.0317 0.0360 0.0399 0.0055 0.0055 0.0708 0.0177 0.0614 AICD † 0.1833 0.1946 0.2393 0.2624 0.1795 0.1606 0.0164 0.2854 0.2656 0.2932 Subclavian Line* 0.0118 0.0119 0.0144 0.0435 0.0080 0.0138 0.0085 0.0231 0.0455 0.0929 Emphysema † 0.2474 0.1001 0.2817 0.2910 0.1197 0.2211 0.0207 0.2952 0.2775 0.3057 Interstitial Edema 0.1930 0.1631 0.2251 0.2186 0.1568 0.1831 0.1004 0.2338 0.1873 0.2288 Hydropneumothorax* 0.0378 0.0273 0.1320 0.0853 0.0163 0.0242 0.0046 0.1074 0.1179 0.1355 Diaphragmatic Hernia † 0.0356 0.0219 0.0812 0.1978 0.0221 0.0277 0.0079 0.1153 0.1669 0.2598 Pericardial Effusion* 0.0246 0.0103 0.0524 0.0612 0.0141 0.0101 0.0041 0.0501 0.0745 0.0557 Healed Rib Fracture † 0.0432 0.0367 0.2191 0.1962 0.0423 0.0407 0.0249 0.2570 0.1632 0.2768 Alveolar Edema † 0.0882 0.0664 0.1057 0.1372 0.0594 0.0825 0.0261 0.1037 0.0830 0.1115 Aortic Valve Replacement* 0.0134 0.0117 0.0381 0.0197 0.0128 0.0143 0.0062 0.0194 0.0225 0.0487 Hilar Enlargement † 0.0312 0.0324 0.1483 0.2044 0.0253 0.0284 0.0199 0.1914 0.1589 0.2078 Internal Jugular Line 0.2703 0.1942 0.3303 0.3378 0.1970 0.2911 0.1191 0.3400 0.3139 0.3526 Normal 0.4369 0.4037 0.5004 0.5040 0.3097 0.4158 0.3408 0.4987 0.4671 0.5068 Diffused Nodules † 0.0708 0.0425 0.2568 0.1905 0.0918 0.0728 0.0099 0.2635 0.1912 0.2303 Calcified Nodules* 0.0130 0.0131 0.0571 0.0258 0.0250 0.0251 0.0114 0.0658 0.1137 0.1205 Clavicle Fracture † 0.0350 0.0227 0.0593 0.0547 0.0205 0.0222 0.0127 0.3349 0.1168 0.1567 Nasogastric Tube 0.4794 0.4292 0.5668 0.5874 0.4207 0.5403 0.3565 0.5842 0.5591 0.6208 Central Venous Line 0.2266 0.1635 0.2802 0.2719 0.1621 0.2319 0.0843 0.2672 0.2693 0.3102 Left Subclavian Line † 0.0269 0.0183 0.0291 0.1051 0.0247 0.0208 0.0101 0.0586 0.1024 0.1373 Pneumomediastinum* 0.0699 0.0230 0.0745 0.0775 0.0215 0.0771 0.0024 0.0833 0.0578 0.0872 Interstitial Thickening † 0.0662 0.0616 0.0808 0.0814 0.0577 0.0761 0.0209 0.0807 0.0754 0.0770 Pulmonary Vascular Congestion 0.1416 0.1119 0.1547 0.1546 0.1108 0.1199 0.0687 0.1527 0.1526 0.1577 Swan Ganz Catheter 0.1584 0.1464 0.3258 0.3693 0.2587 0.2384 0.0665 0.3738 0.4033 0.4627 Aeration* 0.0057 0.0048 0.0058 0.0066 0.0044 0.0039 0.0028 0.0048 0.0095 0.0063 Mediastinal Shift* 0.1320 0.0305 0.2417 0.2226 0.1524 0.0212 0.0057 0.2794 0.1844 0.2347 Surgical Hardware* 0.0097 0.0071 0.0146 0.0059 0.0057 0.0067 0.0076 0.0098 0.0142 0.0136 Patchy Opacities* 0.0218 0.0194 0.0179 0.0190 0.0191 0.0196 0.0130 0.0308 0.0487 0.0383 Mediastinal Clips † 0.0553 0.0569 0.0793 0.0715 0.0480 0.0561 0.0080 0.0875 0.0622 0.1161 Opacity 0.1591 0.1477 0.1958 0.1970 0.1412 0.1504 0.1280 0.1989 0.1961 0.2037 Pulmonary Hypertension* 0.0331 0.0138 0.0296 0.1733 0.0189 0.0315 0.0111 0.1142 0.1154 0.1806 Atelectasis 0.5305 0.5114 0.5909 0.5760 0.4743 0.5329 0.4399 0.5789 0.5585 0.5869 Hyperinflation † 0.0547 0.0442 0.0445 0.0722 0.0452 0.0523 0.0245 0.0751 0.0694 0.0765 ILD 0.3895 0.3390 0.4628 0.4632 0.3100 0.4053 0.1957 0.4647 0.4241 0.4800 Continued on next page 50 Abnormality EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BioViL-T Med-CLIP Med-AR-2B (ours) Med-AR-8B (ours) COPD † 0.0719 0.0344 0.0564 0.0947 0.0334 0.0515 0.0173 0.0824 0.1027 0.0990 Opacification* 0.0311 0.0170 0.0871 0.0586 0.0248 0.0179 0.0069 0.0751 0.0481 0.0651 Mediport* 0.0193 0.0128 0.0612 0.1155 0.0383 0.0121 0.0071 0.1628 0.1989 0.2028 Lung Transplant* 0.0529 0.0310 0.0612 0.0893 0.0470 0.0209 0.0061 0.1473 0.0766 0.1186 Surgical Clips 0.0649 0.0501 0.1679 0.1220 0.0777 0.0453 0.0367 0.1578 0.1907 0.2443 Low Lung Volumes 0.5618 0.5260 0.5758 0.5740 0.4708 0.5296 0.3202 0.5749 0.5605 0.5746 Aortic Aneurysm † 0.0967 0.0184 0.1384 0.2168 0.0608 0.0254 0.0180 0.2240 0.2533 0.2674 Consolidation 0.4609 0.4385 0.5014 0.4946 0.4138 0.4565 0.4068 0.4877 0.4684 0.5020 Pneumonia 0.1883 0.1543 0.2845 0.2805 0.1248 0.2001 0.0913 0.2684 0.2564 0.2780 Free Subdiaphragmatic Air* 0.0203 0.0135 0.0160 0.0650 0.0145 0.0101 0.0077 0.2786 0.1955 0.4584 Fibrosis 0.1383 0.0842 0.2556 0.2903 0.0785 0.1118 0.0485 0.2479 0.2759 0.3226 Lung Mass † 0.0727 0.0601 0.3225 0.3321 0.0528 0.0531 0.0241 0.3169 0.3584 0.3962 Endotracheal Tube 0.5897 0.4998 0.6330 0.6418 0.5398 0.5984 0.4741 0.6557 0.6548 0.6932 Pulmonary Arterial Hypertension Signs* 0.0378 0.0152 0.0692 0.2534 0.0148 0.0482 0.0070 0.1756 0.2089 0.2250 Calcification 0.1994 0.1426 0.2664 0.2827 0.1277 0.1686 0.0695 0.3088 0.2992 0.3630 51 S4.3 SensitivityTable S4: Per-label Sensitivity comparison across baseline, contrastive, and autoregressive models on the CheXpert dataset. Tail abnormalities are marked with an asterisk (*), and medium-prevalence abnormalities are marked with a dagger ( † ). For each row, the best value is shown in bold and the second-best value is underlined. Abnormality EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BioViL-T Med-CLIP Med-AR-2B (ours) Med-AR-8B (ours) LVAD* 0.9038 0.9423 1.0000 0.9615 0.9038 0.8846 0.8269 0.9231 0.9231 0.9808 Diaphragm elevation 0.7598 0.6520 0.7917 0.8284 0.5858 0.6324 0.8701 0.8064 0.8039 0.8260 Osteopenia † 0.8364 0.8121 0.8000 0.9091 0.8182 0.8182 0.7818 0.8727 0.8545 0.9394 Parenchymal opacities † 0.9286 0.8929 0.8036 0.6964 0.7143 0.8929 0.7679 0.8929 0.6964 0.6250 Nodules 0.6770 0.8217 0.7855 0.7726 0.7829 0.6977 0.7054 0.7519 0.6693 0.7933 Pleural effusion 0.8198 0.7939 0.8434 0.8633 0.7847 0.7970 0.8062 0.8266 0.8227 0.8292 Congestive heart failure* 0.8108 0.6216 0.7568 0.7568 0.6757 0.9459 0.8108 0.9189 0.8378 0.6757 Prosthetic valve* 0.9091 0.9394 0.9697 0.8788 0.6061 0.9091 0.5152 0.8485 0.9091 0.9091 Support devices 0.9295 0.8751 0.9285 0.8731 0.9043 0.8711 0.8872 0.9265 0.9053 0.9144 PICC line 0.7285 0.5321 0.9685 0.9382 0.5297 0.6776 0.6194 0.8461 0.9612 0.9721 Pleural thickening 0.6824 0.6318 0.7770 0.7365 0.6655 0.7703 0.7162 0.7973 0.7196 0.8176 Soft tissue swelling † 0.7356 0.7241 0.7586 0.8161 0.7529 0.7011 0.0230 0.7931 0.7586 0.7816 Neoplastic nodules* 0.8837 0.9070 0.9070 0.8837 0.7907 0.7209 0.8605 0.9070 0.7442 0.9767 Nasogastric tube removal* 0.7931 0.5690 0.6207 0.8793 0.6897 0.8793 0.8966 0.7759 0.7586 0.9138 Venous catheter* 0.9286 0.9048 0.7381 0.9524 0.5952 0.8095 0.4286 0.8571 0.8333 0.8095 Hemothorax* 0.6452 0.7097 0.8387 0.7419 0.5484 0.7742 0.6452 0.6774 0.8387 0.7742 CABG † 0.7689 0.8048 0.8446 0.9203 0.7689 0.8486 0.8845 0.8884 0.8167 0.9163 Right subclavian line* 0.7500 0.5750 0.8250 0.6500 0.9000 0.9500 0.6500 0.9000 0.5750 0.8250 Aortic tortuosity † 0.8654 0.8269 0.8782 0.8846 0.8846 0.8654 0.8462 0.8846 0.9423 0.8974 Solitary nodules † 0.7627 0.7797 0.7966 0.8814 0.7797 0.8136 0.8814 0.8305 0.8136 0.7458 Feeding tube removal* 0.4324 0.5946 0.5135 0.7838 0.7297 0.7568 0.8649 0.8108 0.9189 0.7297 Scoliosis † 0.7122 0.6619 0.8129 0.9065 0.7338 0.7626 0.6187 0.8921 0.7338 0.8849 Pulmonary edema 0.8246 0.8096 0.8467 0.8034 0.7309 0.8285 0.7813 0.8261 0.7955 0.8170 Rib fractures 0.6857 0.6571 0.7740 0.6494 0.7065 0.6597 0.6545 0.8156 0.6571 0.7091 ARDS* 0.8462 0.8974 0.8718 0.9231 0.7692 0.8718 0.9231 0.8205 0.8974 0.9231 Aspiration 0.5462 0.6891 0.6597 0.7017 0.7773 0.7647 0.6471 0.6050 0.7143 0.6849 Reticular opacities 0.6761 0.7005 0.7404 0.7506 0.5900 0.7237 0.8715 0.7815 0.7237 0.7198 Kyphosis* 0.9048 0.8095 0.9762 0.9524 0.7381 0.7857 0.7381 0.9286 0.8571 0.9286 Vascular redistribution cephalization † 0.7692 0.7051 0.7949 0.7821 0.6410 0.6154 0.7436 0.7692 0.6026 0.8333 Airspace opacity † 0.5630 0.7059 0.6218 0.6387 0.7815 0.7059 0.8319 0.5042 0.6387 0.6218 Left subclavian catheter* 0.8108 0.9730 0.8649 0.8108 0.7568 0.8919 0.8649 0.8919 0.8919 0.8919 Kerley B lines* 0.7241 0.6897 0.8621 0.7931 0.6207 0.7586 0.6897 0.8966 0.8621 0.9310 Chest tube removal † 0.7273 0.7576 0.7879 0.8636 0.5606 0.6818 0.8788 0.7879 0.7576 0.9394 Degenerative changes † 0.8377 0.8075 0.8151 0.8264 0.7208 0.7962 0.6151 0.8566 0.8453 0.8642 Dialysis catheter* 0.6857 0.8571 0.8000 0.7143 0.7143 0.4857 0.8000 0.7714 0.7714 0.8857 Hilar lymphadenopathy* 0.8400 0.6800 0.9200 0.6200 0.6200 0.8600 0.8600 0.8000 0.5600 0.8200 Pigtail catheter † 0.8611 0.6944 0.8889 0.7778 0.7222 0.9583 0.9028 0.8472 0.8194 0.9722 Epidural catheter † 0.9271 0.8646 0.9479 0.9792 0.8229 0.8229 0.9583 0.9792 0.9583 0.9792 Mediastinal drain † 0.8571 0.8631 0.9524 0.8929 0.8988 0.9107 0.8750 0.9107 0.8929 0.9226 Retrocardiac opacity 0.7765 0.8314 0.7330 0.8580 0.6989 0.8371 0.8769 0.7784 0.7765 0.7879 Pneumothorax 0.7542 0.7375 0.8904 0.8228 0.6146 0.7508 0.7442 0.8494 0.8283 0.8704 Prosthetic aortic valve* 0.9111 0.9111 0.9556 0.8222 0.7778 0.9111 0.6667 0.9111 0.8000 0.9556 Extubation † 0.7459 0.9098 0.7869 0.7787 0.8443 0.7623 0.8607 0.7869 0.8525 0.8934 Perihilar opacities 0.5870 0.6196 0.7754 0.7391 0.6739 0.7029 0.8370 0.8007 0.7536 0.7572 Sternotomy wires 0.9681 0.9389 0.9806 0.9778 0.7889 0.9708 0.6861 0.9722 0.9750 0.9944 Continued on next page 52 Abnormality EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BioViL-T Med-CLIP Med-AR-2B (ours) Med-AR-8B (ours) Subcutaneous emphysema † 0.8605 0.7364 0.8915 0.8760 0.6279 0.7984 0.3721 0.8837 0.9225 0.9302 MPA enlargement † 0.6234 0.8701 0.8182 0.8312 0.6234 0.8052 0.5584 0.7532 0.7532 0.8052 Bronchial wall thickening* 0.7407 0.7593 0.7407 0.7593 0.7037 0.7778 0.7222 0.7963 0.8333 0.8148 Chest wall deformity † 0.6991 0.3982 0.7611 0.8053 0.5487 0.5487 0.6549 0.7611 0.4956 0.6903 Feeding tube 0.9492 0.9383 0.9492 0.9520 0.8765 0.9259 0.8999 0.9643 0.9438 0.9492 Tracheostomy cannula* 0.8684 0.9211 1.0000 1.0000 0.6053 0.8421 0.9474 1.0000 1.0000 1.0000 Sternotomy † 0.9760 0.9520 0.9760 0.9920 0.7040 0.9360 0.1200 0.9840 0.9760 0.9920 Granuloma † 0.6849 0.6781 0.7877 0.7397 0.8014 0.7466 0.7740 0.8767 0.7192 0.8219 Mediastinal mass* 0.5636 0.9091 0.8727 0.8000 0.6909 0.8364 0.8909 0.6727 0.7455 0.7273 Basilar opacity 0.7560 0.7404 0.7986 0.7957 0.6383 0.6809 0.7858 0.7631 0.6979 0.8255 Artifacts 0.5829 0.5829 0.6326 0.6000 0.4512 0.4264 0.8899 0.6186 0.6326 0.5659 Cardiomegaly 0.7946 0.7381 0.8269 0.8011 0.7158 0.7346 0.6553 0.8363 0.8145 0.8239 Solitary pulmonary nodule † 0.6824 0.8118 0.7647 0.7529 0.6941 0.7647 0.7176 0.7765 0.7647 0.7765 Mediastinal widening 0.7311 0.7180 0.7115 0.7705 0.6852 0.6918 0.9148 0.7180 0.7049 0.8000 Improved aeration* 0.5676 0.5541 0.5541 0.6892 0.6892 0.8378 0.6081 0.9459 0.7027 0.7973 Chest tube 0.8581 0.8431 0.8990 0.8827 0.7763 0.8226 0.7681 0.8772 0.8977 0.9209 Malignant nodules* 0.9091 0.8409 0.9318 0.8409 0.8864 0.7955 0.8182 0.8636 0.8182 0.9091 Tracheostomy tube † 0.9070 0.8605 0.9734 0.9701 0.7774 0.8904 0.7674 0.9568 0.9767 0.9834 Cancer 0.6367 0.6609 0.6782 0.7232 0.7163 0.7232 0.5433 0.7924 0.7163 0.7958 Pacemaker † 0.9688 0.9107 0.9732 0.9732 0.8661 0.9152 0.9241 0.9866 0.9464 0.9955 Mediastinal drains † 0.9878 0.9390 0.9756 1.0000 0.8537 0.9634 0.8049 0.9878 0.9878 0.9756 Atrial dilation* 0.6190 0.7143 0.7857 0.8571 0.8095 0.6190 0.5952 0.8333 0.8095 0.6905 AICD † 0.9762 0.9762 0.9921 1.0000 0.9365 0.9603 0.5952 0.9921 0.9603 0.9921 Subclavian line* 0.6604 0.8113 0.8302 0.8113 0.8868 0.7925 0.6981 0.7547 0.7925 0.8679 Emphysema † 0.8176 0.8176 0.8706 0.8706 0.6471 0.8176 0.5706 0.9059 0.8824 0.8765 Interstitial edema 0.7993 0.8017 0.7433 0.7786 0.6910 0.8200 0.8224 0.8187 0.8418 0.8504 Hydropneumothorax* 0.8788 0.8182 0.8485 0.8788 0.7576 0.9091 0.8788 0.9394 0.8485 0.8788 Diaphragmatic hernia † 0.7246 0.6087 0.7246 0.8986 0.6812 0.8116 0.5652 0.9130 0.8551 0.7971 Pericardial effusion* 0.7872 0.7447 0.8511 0.7660 0.7660 0.7447 0.8511 0.9149 0.8936 0.6596 Healed rib fracture † 0.7037 0.6914 0.7346 0.8148 0.6543 0.6852 0.6543 0.8889 0.7222 0.7284 Alveolar edema † 0.7699 0.7345 0.7699 0.8673 0.6549 0.7876 0.8319 0.7876 0.7699 0.8584 Aortic valve replacement* 0.9767 0.9070 0.8372 0.9535 0.7907 0.9070 0.6744 0.9767 0.8837 0.9302 Hilar enlargement † 0.6807 0.5422 0.7530 0.7771 0.8193 0.7169 0.7530 0.7892 0.7169 0.8855 Internal jugular line 0.8755 0.9098 0.9047 0.8895 0.7903 0.8793 0.7459 0.8831 0.8907 0.9187 Normal 0.8986 0.8784 0.8953 0.8767 0.8649 0.8953 0.8311 0.8378 0.8547 0.8750 Diffused nodules † 0.7667 0.6667 0.8667 0.8500 0.6667 0.7833 0.7000 0.8500 0.8500 0.9167 Calcified nodules* 0.7250 0.7750 0.8750 0.6500 0.6250 0.9500 0.6000 0.5250 0.5750 0.8250 Clavicle fracture † 0.4773 0.7045 0.7273 0.7045 0.5682 0.5341 0.7159 0.8636 0.5795 0.7727 Nasogastric tube 0.9427 0.9379 0.9510 0.9379 0.9032 0.9450 0.9259 0.9438 0.9223 0.9450 Central venous line 0.8281 0.8790 0.8597 0.8790 0.8047 0.8886 0.7056 0.8776 0.8680 0.8803 Left subclavian line † 0.8533 0.7067 0.8933 0.7600 0.7067 0.6800 0.8133 0.7067 0.9200 0.8533 Pneumomediastinum* 0.7500 0.6250 0.7812 0.7500 0.6875 0.7188 0.9688 0.7188 0.6562 0.7812 Interstitial thickening † 0.6601 0.6601 0.7044 0.6305 0.6256 0.6995 0.3153 0.7241 0.5961 0.6650 Pulmonary vascular congestion 0.7268 0.5176 0.6773 0.6789 0.5240 0.6054 0.8722 0.6374 0.5479 0.7077 Swan-Ganz catheter 0.8963 0.8506 0.9212 0.8880 0.8174 0.8548 0.8797 0.8714 0.8797 0.8963 Aeration* 0.6486 0.7027 0.7297 0.8919 0.8649 0.6486 1.0000 0.6757 0.8378 0.6486 Mediastinal shift* 0.9123 0.6667 0.8596 0.8947 0.7895 0.8772 0.5614 0.9123 0.8596 0.8596 Surgical hardware* 0.8409 0.5000 0.9091 0.6591 0.7727 0.5227 0.1136 0.6364 0.7727 0.8636 Patchy opacities* 0.7778 0.6190 0.8095 0.7302 0.6190 0.8571 0.6984 0.7619 0.8095 0.7143 Mediastinal clips † 0.9053 0.8737 0.9263 0.9053 0.7789 0.8211 0.9895 0.8316 0.8947 0.9158 Opacity 0.8458 0.8153 0.8300 0.7702 0.7775 0.8017 0.8269 0.7828 0.7062 0.6957 Pulmonary hypertension* 0.6964 0.5179 0.6964 0.7321 0.4821 0.6071 0.2321 0.7679 0.6429 0.6786 Atelectasis 0.7538 0.7956 0.7871 0.7714 0.7683 0.8101 0.8093 0.7386 0.7703 0.7120 Hyperinflation † 0.4926 0.5735 0.5662 0.6324 0.4191 0.4338 0.6765 0.6691 0.4706 0.7941 Continued on next page 53 Abnormality EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BioViL-T Med-CLIP Med-AR-2B (ours) Med-AR-8B (ours) ILD 0.6978 0.6029 0.7652 0.7707 0.6105 0.6374 0.6780 0.6341 0.7235 0.7888 COPD † 0.7021 0.6489 0.7021 0.7660 0.7340 0.7234 0.5532 0.7553 0.8191 0.7128 Opacification* 0.8056 0.6389 0.7500 0.7500 0.7778 0.6389 0.8611 0.6944 0.8056 0.8056 Mediport* 0.8696 0.8043 0.9130 0.8261 0.6087 0.8478 0.3696 0.9348 0.9565 1.0000 Lung transplant* 0.8475 0.7458 0.9322 0.8983 0.6780 0.8475 0.6441 0.9322 0.8814 0.9492 Surgical clips 0.6688 0.6369 0.7325 0.7261 0.7102 0.7389 0.5223 0.7516 0.7771 0.8662 Low lung volumes 0.7502 0.6973 0.7393 0.7507 0.7364 0.7235 0.7007 0.7598 0.7459 0.7869 Aortic aneurysm † 0.7535 0.5211 0.7254 0.8310 0.6338 0.6549 0.5915 0.8732 0.7887 0.8451 Consolidation 0.7867 0.8050 0.8605 0.8512 0.7791 0.8605 0.7601 0.8875 0.8177 0.8367 Pneumonia 0.5973 0.6183 0.7720 0.6877 0.6481 0.6010 0.6444 0.6196 0.6927 0.7076 Free subdiaphragmatic air* 0.7045 0.7500 0.6591 0.7727 0.5682 0.6364 0.8182 0.8864 0.7045 0.7955 Fibrosis 0.8339 0.7934 0.8081 0.7675 0.7380 0.7860 0.6384 0.7749 0.8266 0.8266 Lung mass † 0.8000 0.7150 0.7650 0.7800 0.6750 0.7400 0.8300 0.8100 0.8050 0.7750 Endotracheal tube 0.9515 0.9451 0.9570 0.9634 0.9121 0.9542 0.9011 0.9542 0.9460 0.9643 Pulmonary arterial hypertension signs* 0.7872 0.7447 0.7021 0.7447 0.6383 0.8511 0.7872 0.8298 0.6596 0.7447 Calcification 0.6925 0.6789 0.7660 0.8143 0.6228 0.7389 0.6905 0.8414 0.7118 0.7795 54 S4.4 SpecificityTable S5: Per-label Specificity comparison across baseline, contrastive, and autoregressive models on the CheXpert dataset. Tail abnormalities are marked with an asterisk (*), and medium-prevalence abnormalities are marked with a dagger ( † ). For each row, the best value is shown in bold and the second-best value is underlined. Abnormality EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BioViL-T Med-CLIP Med-AR-2B (ours) Med-AR-8B (ours) LVAD* 0.8065 0.7863 0.8010 0.9075 0.7148 0.8121 0.4776 0.8730 0.8487 0.9339 Diaphragm elevation 0.6971 0.6838 0.7580 0.8405 0.7724 0.7352 0.2505 0.8470 0.8271 0.8646 Osteopenia † 0.8344 0.7647 0.8530 0.7553 0.7418 0.7813 0.5046 0.8146 0.8315 0.7632 Parenchymal opacities † 0.4206 0.5634 0.6590 0.6801 0.5716 0.5945 0.6355 0.5439 0.6716 0.8011 Nodules 0.7293 0.5739 0.7430 0.7044 0.6098 0.6963 0.6669 0.7201 0.7802 0.7214 Pleural effusion 0.7838 0.7552 0.8211 0.8014 0.7058 0.8035 0.6161 0.8244 0.8040 0.8214 Congestive heart failure* 0.6798 0.8383 0.8273 0.8566 0.7916 0.5695 0.3998 0.6965 0.8279 0.8973 Prosthetic valve* 0.7597 0.7262 0.7028 0.7465 0.7734 0.7253 0.6287 0.7924 0.6649 0.7571 Support devices 0.6490 0.6782 0.6661 0.7194 0.6327 0.6948 0.6317 0.6662 0.6807 0.6859 PICC line 0.5796 0.6904 0.8667 0.8760 0.6741 0.5623 0.4518 0.8013 0.8517 0.8774 Pleural thickening 0.7402 0.7549 0.7308 0.7794 0.7030 0.6493 0.5405 0.7405 0.7744 0.7502 Soft tissue swelling † 0.9070 0.7672 0.9634 0.9180 0.6804 0.8983 0.9921 0.9304 0.9348 0.9519 Neoplastic nodules* 0.6885 0.6321 0.9259 0.9378 0.7301 0.7914 0.4662 0.8758 0.9235 0.8724 Nasogastric tube removal* 0.6783 0.8275 0.8460 0.6569 0.7104 0.4541 0.3930 0.7936 0.7488 0.6156 Venous catheter* 0.4767 0.4763 0.6425 0.5309 0.6982 0.6031 0.8085 0.6680 0.7358 0.7646 Hemothorax* 0.8127 0.8055 0.8013 0.7962 0.8324 0.6726 0.6389 0.9489 0.7114 0.8283 CABG † 0.7182 0.6645 0.6577 0.6168 0.5686 0.5678 0.1737 0.6337 0.7151 0.6228 Right subclavian line* 0.6601 0.8152 0.6344 0.8180 0.4388 0.3950 0.6834 0.6948 0.9265 0.9068 Aortic tortuosity † 0.7795 0.7671 0.8202 0.8397 0.6854 0.7539 0.5704 0.8356 0.8014 0.8777 Solitary nodules † 0.7465 0.7243 0.7853 0.6724 0.6935 0.6792 0.6155 0.7817 0.7677 0.8530 Feeding tube removal* 0.8434 0.5565 0.7502 0.5867 0.5559 0.4808 0.3464 0.5548 0.3149 0.7093 Scoliosis † 0.7892 0.6971 0.6932 0.8307 0.6453 0.6458 0.5724 0.7513 0.8364 0.8523 Pulmonary edema 0.7168 0.6798 0.7294 0.7703 0.6973 0.7019 0.6140 0.7464 0.7388 0.7611 Rib fractures 0.6362 0.5897 0.7455 0.8426 0.5378 0.5906 0.4469 0.6780 0.7707 0.8128 ARDS* 0.8446 0.8170 0.8683 0.8169 0.8425 0.8491 0.7314 0.8739 0.8272 0.8412 Aspiration 0.7653 0.5543 0.7452 0.7302 0.4375 0.5446 0.5100 0.8064 0.6720 0.7516 Reticular opacities 0.7253 0.6601 0.7278 0.6957 0.6625 0.6927 0.2024 0.6921 0.7084 0.7664 Kyphosis* 0.7583 0.7269 0.7287 0.8735 0.8058 0.7733 0.6532 0.8515 0.9099 0.8698 Vascular redistribution cephalization † 0.7451 0.7558 0.7383 0.7385 0.7566 0.8089 0.4220 0.7904 0.8584 0.7316 Airspace opacity † 0.7661 0.6309 0.7609 0.7413 0.4947 0.6937 0.4445 0.8446 0.6641 0.7325 Left subclavian catheter* 0.4734 0.1708 0.6217 0.7688 0.5558 0.2416 0.3693 0.6429 0.7816 0.8745 Kerley B lines* 0.8742 0.8324 0.8148 0.8643 0.8156 0.8023 0.4513 0.7841 0.7745 0.7431 Chest tube removal † 0.7876 0.6967 0.7505 0.7128 0.8037 0.7906 0.2375 0.7582 0.7776 0.6446 Degenerative changes † 0.7000 0.6771 0.7302 0.7656 0.7319 0.7076 0.7219 0.7414 0.7325 0.7396 Dialysis catheter* 0.7991 0.5263 0.8202 0.9060 0.8054 0.7835 0.3845 0.9098 0.9185 0.9522 Hilar lymphadenopathy* 0.5909 0.7287 0.6084 0.8517 0.6984 0.5329 0.3956 0.8016 0.8874 0.8115 Pigtail catheter † 0.5490 0.6363 0.9105 0.8178 0.6320 0.4139 0.2523 0.8570 0.9484 0.9227 Epidural catheter † 0.8202 0.7651 0.8200 0.8574 0.8134 0.7800 0.2207 0.9231 0.9465 0.9750 Mediastinal drain † 0.8916 0.8455 0.8580 0.9053 0.7089 0.8141 0.4795 0.8627 0.8667 0.9021 Retrocardiac opacity 0.6092 0.4950 0.6923 0.5779 0.5984 0.5298 0.4080 0.6686 0.6438 0.6835 Pneumothorax 0.7697 0.7544 0.8526 0.8544 0.7625 0.8051 0.4877 0.8297 0.8460 0.8552 Prosthetic aortic valve* 0.5757 0.7263 0.6933 0.7898 0.7438 0.6814 0.5739 0.7157 0.8080 0.7178 Extubation † 0.7426 0.5608 0.7274 0.7249 0.5997 0.6610 0.5167 0.8021 0.7451 0.7457 Perihilar opacities 0.8010 0.7425 0.7525 0.7954 0.6191 0.7351 0.4240 0.7282 0.6982 0.7849 Sternotomy wires 0.7642 0.7754 0.7612 0.7577 0.7445 0.7528 0.4391 0.7685 0.7517 0.7591 Continued on next page 55 Abnormality EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BioViL-T Med-CLIP Med-AR-2B (ours) Med-AR-8B (ours) Subcutaneous emphysema † 0.8526 0.7636 0.9098 0.9412 0.8285 0.8836 0.7480 0.9234 0.8616 0.9454 MPA enlargement † 0.8327 0.4813 0.6594 0.8117 0.6603 0.5971 0.5937 0.9371 0.8398 0.8601 Bronchial wall thickening* 0.7789 0.6975 0.7970 0.7784 0.7113 0.6877 0.6282 0.7581 0.6342 0.7116 Chest wall deformity † 0.6428 0.8089 0.6360 0.6533 0.6564 0.7008 0.5331 0.7077 0.8730 0.8098 Feeding tube 0.8259 0.7486 0.8725 0.8890 0.7633 0.8512 0.6616 0.8726 0.8812 0.8948 Tracheostomy cannula* 0.8286 0.6032 0.9437 0.9405 0.8979 0.7213 0.4341 0.9360 0.9235 0.9395 Sternotomy † 0.7158 0.7010 0.7128 0.7053 0.7587 0.7428 0.9030 0.7167 0.6973 0.7082 Granuloma † 0.7240 0.6968 0.7822 0.7398 0.5717 0.6808 0.5912 0.7324 0.8631 0.8645 Mediastinal mass* 0.7400 0.4473 0.6864 0.7792 0.5139 0.5470 0.3978 0.9002 0.8525 0.9117 Basilar opacity 0.6003 0.5893 0.5854 0.6036 0.6610 0.6597 0.4999 0.6244 0.6890 0.5918 Artifacts 0.6290 0.5485 0.6167 0.6882 0.7141 0.7366 0.1523 0.6726 0.6202 0.7254 Cardiomegaly 0.7952 0.7428 0.7992 0.8218 0.7093 0.7543 0.5011 0.7832 0.7882 0.8098 Solitary pulmonary nodule † 0.7722 0.6357 0.6889 0.6799 0.7115 0.6851 0.7197 0.6946 0.7543 0.7317 Mediastinal widening 0.7205 0.5730 0.8095 0.7921 0.6163 0.6192 0.1523 0.8481 0.8385 0.8183 Improved aeration* 0.6118 0.5779 0.6247 0.5429 0.5402 0.2809 0.4731 0.2171 0.4628 0.4115 Chest tube 0.8512 0.7832 0.9020 0.8909 0.7421 0.8107 0.4706 0.8869 0.8866 0.8988 Malignant nodules* 0.6569 0.7051 0.9060 0.9330 0.6749 0.7304 0.5307 0.9086 0.8253 0.9250 Tracheostomy tube † 0.9118 0.6905 0.9474 0.9585 0.8657 0.8932 0.6146 0.9523 0.9584 0.9586 Cancer 0.7815 0.7272 0.8404 0.8191 0.6066 0.6512 0.6350 0.7342 0.8087 0.7824 Pacemaker † 0.9221 0.8609 0.9135 0.9183 0.8400 0.8662 0.1672 0.9124 0.9321 0.9186 Mediastinal drains † 0.8409 0.8610 0.8840 0.8549 0.8400 0.8295 0.5867 0.8726 0.8534 0.9011 Atrial dilation* 0.8044 0.7142 0.6975 0.5688 0.5898 0.6350 0.6170 0.6445 0.5984 0.8311 AICD † 0.9136 0.8763 0.9101 0.9200 0.8936 0.8819 0.5800 0.9314 0.9539 0.9442 Subclavian line* 0.6946 0.4753 0.5698 0.7387 0.3108 0.5005 0.4785 0.8225 0.8685 0.8983 Emphysema † 0.8409 0.7178 0.8382 0.8752 0.8377 0.8343 0.5991 0.8362 0.8432 0.8916 Interstitial edema 0.5881 0.5421 0.7062 0.6316 0.5899 0.5669 0.3719 0.6210 0.5465 0.6034 Hydropneumothorax* 0.8063 0.6944 0.9011 0.8719 0.6173 0.7671 0.3999 0.8759 0.8666 0.9313 Diaphragmatic hernia † 0.8058 0.8104 0.8465 0.7333 0.8048 0.6374 0.6308 0.7235 0.7388 0.8903 Pericardial effusion* 0.6910 0.6259 0.7084 0.8006 0.5906 0.5966 0.2708 0.6865 0.6126 0.8913 Healed rib fracture † 0.6677 0.6598 0.8290 0.7378 0.6833 0.6865 0.6159 0.7091 0.6910 0.8809 Alveolar edema † 0.8285 0.8146 0.8758 0.7556 0.8040 0.8070 0.5174 0.8776 0.8219 0.7822 Aortic valve replacement* 0.6105 0.7102 0.8308 0.7318 0.6917 0.6688 0.5911 0.7012 0.7344 0.7573 Hilar enlargement † 0.6532 0.6855 0.7521 0.8114 0.3657 0.5532 0.4419 0.8022 0.8300 0.7484 Internal jugular line 0.6663 0.5834 0.7217 0.7678 0.6319 0.7057 0.5470 0.7874 0.7595 0.7649 Normal 0.7937 0.8057 0.8246 0.8404 0.7576 0.7875 0.7914 0.8684 0.8581 0.8355 Diffused nodules † 0.8153 0.8184 0.9356 0.9111 0.8814 0.7848 0.5928 0.8964 0.8585 0.9264 Calcified nodules* 0.7297 0.6216 0.6766 0.7949 0.7683 0.4964 0.8102 0.9604 0.9537 0.7925 Clavicle fracture † 0.8854 0.6844 0.7350 0.7152 0.7078 0.7908 0.4775 0.8552 0.8156 0.7046 Nasogastric tube 0.8317 0.8299 0.8459 0.8608 0.7922 0.8395 0.7711 0.8622 0.8562 0.8836 Central venous line 0.5982 0.5058 0.6815 0.6685 0.5319 0.5351 0.3999 0.6452 0.6979 0.6982 Left subclavian line † 0.6147 0.6811 0.6579 0.8906 0.6082 0.7288 0.4205 0.8576 0.8087 0.9373 Pneumomediastinum* 0.7765 0.7100 0.8040 0.8287 0.7643 0.7372 0.0531 0.7984 0.8917 0.8815 Interstitial thickening † 0.7194 0.7007 0.7057 0.7428 0.6452 0.7137 0.7344 0.6840 0.7403 0.7368 Pulmonary vascular congestion 0.5896 0.7264 0.6893 0.6661 0.7064 0.6850 0.2695 0.7017 0.7505 0.6551 Swan-Ganz catheter 0.7351 0.7623 0.9009 0.8984 0.8793 0.8314 0.5826 0.9354 0.9412 0.9616 Aeration* 0.5582 0.5495 0.4449 0.2714 0.2480 0.4667 0.0363 0.4840 0.4088 0.5875 Mediastinal shift* 0.7997 0.8492 0.8742 0.9279 0.8394 0.6985 0.5675 0.9083 0.9043 0.9490 Surgical hardware* 0.4012 0.7686 0.4675 0.6127 0.3991 0.7519 0.9651 0.6801 0.6301 0.6247 Patchy opacities* 0.6689 0.7197 0.6281 0.7479 0.6630 0.6015 0.5978 0.7616 0.6786 0.7924 Mediastinal clips † 0.6898 0.6860 0.7388 0.7641 0.7113 0.7590 0.0448 0.8458 0.7471 0.7659 Opacity 0.4543 0.4634 0.5143 0.5630 0.4454 0.4786 0.4239 0.5425 0.5980 0.6371 Pulmonary hypertension* 0.6980 0.7732 0.7195 0.8644 0.7788 0.6473 0.8667 0.8589 0.8854 0.9276 Atelectasis 0.5463 0.4786 0.5682 0.5831 0.4513 0.4945 0.4035 0.6055 0.5440 0.6345 Hyperinflation † 0.8350 0.7258 0.7801 0.7453 0.8513 0.8733 0.5303 0.6693 0.8446 0.5702 Continued on next page 56 Abnormality EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BioViL-T Med-CLIP Med-AR-2B (ours) Med-AR-8B (ours) ILD 0.6579 0.6939 0.6456 0.6320 0.6185 0.7132 0.4150 0.7659 0.6394 0.6329 COPD † 0.7585 0.8004 0.8197 0.7621 0.7141 0.7224 0.6692 0.7817 0.6714 0.8422 Opacification* 0.8094 0.8604 0.8974 0.7688 0.7674 0.7775 0.4240 0.8799 0.6396 0.8291 Mediport* 0.7025 0.6868 0.8857 0.9104 0.8069 0.4599 0.7368 0.8622 0.8901 0.9598 Lung transplant* 0.7771 0.7435 0.7977 0.8615 0.7748 0.7125 0.4123 0.7932 0.9023 0.8539 Surgical clips 0.6295 0.5696 0.7546 0.7034 0.5393 0.4520 0.6012 0.7146 0.7690 0.7884 Low lung volumes 0.7645 0.7774 0.7868 0.7752 0.6811 0.7501 0.5812 0.7775 0.7577 0.7394 Aortic aneurysm † 0.6151 0.6450 0.7409 0.7089 0.6658 0.5545 0.5781 0.6964 0.7987 0.8165 Consolidation 0.5772 0.5457 0.5459 0.5432 0.5304 0.5047 0.5379 0.5008 0.5379 0.5521 Pneumonia 0.7436 0.6468 0.6568 0.7499 0.5618 0.7408 0.4655 0.7892 0.6884 0.7143 Free subdiaphragmatic air* 0.6302 0.5687 0.7054 0.7401 0.6842 0.6607 0.4325 0.7931 0.8482 0.9394 Fibrosis 0.7287 0.7280 0.8330 0.8715 0.7100 0.7512 0.7046 0.8751 0.7791 0.8414 Lung mass † 0.6917 0.6769 0.8677 0.8537 0.6884 0.6769 0.4062 0.8523 0.7939 0.8994 Endotracheal tube 0.8670 0.8325 0.8879 0.8884 0.8286 0.8566 0.8281 0.8951 0.8848 0.9010 Pulmonary arterial hypertension signs* 0.7279 0.6811 0.8046 0.9328 0.6914 0.4807 0.4189 0.8855 0.8825 0.9045 Calcification 0.7873 0.7345 0.7624 0.7519 0.7473 0.7064 0.5367 0.7361 0.8068 0.8350 57 S4.5 Youden’s J Table S6: Per-label Youden’s J Score comparison across baseline, contrastive, and autoregressive models on the CheXpert dataset. Tail abnormalities are marked with an asterisk (*), and medium-prevalence abnormalities are marked with a dagger ( † ). For each row, the best value is shown in bold and the second-best value is underlined. Abnormality EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BioViL-T Med-CLIP Med-AR-2B (ours) Med-AR-8B (ours) LVAD* 0.7103 0.7286 0.8010 0.8690 0.6186 0.6967 0.3045 0.7961 0.7718 0.9147 Diaphragm elevation 0.4569 0.3358 0.5497 0.6689 0.3582 0.3676 0.1206 0.6534 0.6310 0.6906 Osteopenia † 0.6708 0.5768 0.6530 0.6644 0.5600 0.5995 0.2864 0.6873 0.6860 0.7026 Parenchymal opacities † 0.3492 0.4563 0.4626 0.3765 0.2859 0.4874 0.4034 0.4368 0.3680 0.4261 Nodules 0.4063 0.3956 0.5285 0.4770 0.3927 0.3940 0.3723 0.4720 0.4495 0.5147 Pleural effusion 0.6036 0.5491 0.6645 0.6647 0.4905 0.6005 0.4223 0.6510 0.6267 0.6506 Congestive heart failure* 0.4906 0.4599 0.5841 0.6134 0.4673 0.5154 0.2106 0.6154 0.6657 0.5730 Prosthetic valve* 0.6688 0.6656 0.6725 0.6253 0.3795 0.6344 0.1439 0.6409 0.5740 0.6662 Support devices 0.5785 0.5533 0.5946 0.5925 0.5370 0.5659 0.5189 0.5927 0.5860 0.6003 PICC line 0.3081 0.2225 0.8352 0.8142 0.2038 0.2399 0.0712 0.6474 0.8129 0.8495 Pleural thickening 0.4226 0.3867 0.5078 0.5159 0.3685 0.4196 0.2567 0.5378 0.4940 0.5678 Soft tissue swelling † 0.6426 0.4913 0.7220 0.7341 0.4333 0.5994 0.0151 0.7235 0.6934 0.7335 Neoplastic nodules* 0.5722 0.5391 0.8329 0.8215 0.5208 0.5123 0.3267 0.7828 0.6677 0.8491 Nasogastric tube removal* 0.4714 0.3965 0.4667 0.5362 0.4001 0.3334 0.2896 0.5695 0.5074 0.5294 Venous catheter* 0.4053 0.3811 0.3806 0.4833 0.2934 0.4126 0.2371 0.5251 0.5691 0.5741 Hemothorax* 0.4579 0.5152 0.6400 0.5381 0.3808 0.4468 0.2841 0.6263 0.5501 0.6025 CABG † 0.4871 0.4693 0.5023 0.5371 0.3375 0.4164 0.0582 0.5221 0.5318 0.5391 Right subclavian line* 0.4101 0.3902 0.4594 0.4680 0.3388 0.3450 0.3334 0.5948 0.5015 0.7318 Aortic tortuosity † 0.6449 0.5940 0.6984 0.7243 0.5700 0.6193 0.4166 0.7202 0.7437 0.7751 Solitary nodules † 0.5092 0.5040 0.5819 0.5538 0.4732 0.4928 0.4969 0.6122 0.5813 0.5988 Feeding tube removal* 0.2758 0.1511 0.2637 0.3705 0.2856 0.2376 0.2113 0.3656 0.2338 0.4390 Scoliosis † 0.5014 0.3590 0.5061 0.7372 0.3791 0.4084 0.1911 0.6434 0.5702 0.7372 Pulmonary edema 0.5414 0.4894 0.5761 0.5737 0.4282 0.5304 0.3953 0.5725 0.5343 0.5781 Rib fractures 0.3219 0.2468 0.5195 0.4920 0.2443 0.2503 0.1014 0.4936 0.4278 0.5219 ARDS* 0.6908 0.7144 0.7401 0.7400 0.6117 0.7209 0.6545 0.6944 0.7246 0.7643 Aspiration 0.3115 0.2434 0.4049 0.4319 0.2148 0.3093 0.1571 0.4114 0.3863 0.4365 Reticular opacities 0.4014 0.3606 0.4682 0.4463 0.2525 0.4164 0.0739 0.4736 0.4321 0.4862 Kyphosis* 0.6631 0.5364 0.7049 0.8259 0.5439 0.5590 0.3913 0.7801 0.7670 0.7984 Vascular redistribution cephalization † 0.5143 0.4609 0.5332 0.5206 0.3976 0.4243 0.1656 0.5596 0.4610 0.5649 Airspace opacity † 0.3291 0.3368 0.3827 0.3800 0.2762 0.3996 0.2764 0.3488 0.3028 0.3543 Left subclavian catheter* 0.2842 0.1438 0.4866 0.5796 0.3126 0.1335 0.2342 0.5348 0.6735 0.7664 Kerley B lines* 0.5983 0.5221 0.6769 0.6574 0.4363 0.5609 0.1410 0.6807 0.6366 0.6741 Chest tube removal † 0.5149 0.4543 0.5384 0.5764 0.3643 0.4724 0.1163 0.5461 0.5352 0.5840 Degenerative changes † 0.5377 0.4846 0.5453 0.5920 0.4527 0.5038 0.3370 0.5980 0.5778 0.6038 Dialysis catheter* 0.4848 0.3834 0.6202 0.6203 0.5197 0.2692 0.1845 0.6812 0.6899 0.8379 Hilar lymphadenopathy* 0.4309 0.4087 0.5284 0.4717 0.3184 0.3929 0.2556 0.6016 0.4474 0.6315 Pigtail catheter † 0.4101 0.3307 0.7994 0.5956 0.3542 0.3722 0.1551 0.7042 0.7678 0.8949 Epidural catheter † 0.7473 0.6297 0.7679 0.8366 0.6363 0.6029 0.1790 0.9023 0.9048 0.9542 Mediastinal drain † 0.7487 0.7086 0.8104 0.7982 0.6077 0.7248 0.3545 0.7734 0.7596 0.8247 Retrocardiac opacity 0.3857 0.3264 0.4253 0.4359 0.2973 0.3669 0.2849 0.4470 0.4203 0.4714 Pneumothorax 0.5239 0.4919 0.7430 0.6772 0.3771 0.5559 0.2319 0.6791 0.6743 0.7256 Prosthetic aortic valve* 0.4868 0.6374 0.6489 0.6120 0.5216 0.5925 0.2406 0.6268 0.6080 0.6734 Extubation † 0.4885 0.4706 0.5143 0.5036 0.4440 0.4233 0.3774 0.5890 0.5976 0.6391 Perihilar opacities 0.3880 0.3621 0.5279 0.5345 0.2930 0.4380 0.2610 0.5289 0.4518 0.5421 Sternotomy wires 0.7323 0.7143 0.7418 0.7355 0.5334 0.7236 0.1252 0.7407 0.7267 0.7535 Subcutaneous emphysema † 0.7131 0.5000 0.8013 0.8172 0.4564 0.6820 0.1201 0.8071 0.7841 0.8756 Continued on next page 58 Abnormality EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BioViL-T Med-CLIP Med-AR-2B (ours) Med-AR-8B (ours) MPA enlargement † 0.4561 0.3514 0.4776 0.6429 0.2837 0.4023 0.1521 0.6903 0.5930 0.6653 Bronchial wall thickening* 0.5196 0.4568 0.5377 0.5377 0.4150 0.4655 0.3504 0.5544 0.4675 0.5264 Chest wall deformity † 0.3419 0.2071 0.3971 0.4586 0.2051 0.2495 0.1880 0.4688 0.3686 0.5001 Feeding tube 0.7751 0.6869 0.8217 0.8410 0.6398 0.7771 0.5615 0.8369 0.8250 0.8440 Tracheostomy cannula* 0.6970 0.5243 0.9437 0.9405 0.5032 0.5634 0.3815 0.9360 0.9235 0.9395 Sternotomy † 0.6918 0.6530 0.6888 0.6973 0.4627 0.6788 0.0230 0.7007 0.6733 0.7002 Granuloma † 0.4089 0.3749 0.5699 0.4795 0.3731 0.4274 0.3652 0.6091 0.5823 0.6864 Mediastinal mass* 0.3036 0.3564 0.5591 0.5792 0.2048 0.3834 0.2887 0.5729 0.5980 0.6390 Basilar opacity 0.3563 0.3297 0.3840 0.3993 0.2993 0.3406 0.2857 0.3875 0.3869 0.4173 Artifacts 0.2119 0.1314 0.2493 0.2882 0.1653 0.1630 0.0422 0.2912 0.2528 0.2913 Cardiomegaly 0.5898 0.4809 0.6261 0.6229 0.4251 0.4889 0.1564 0.6195 0.6027 0.6337 Solitary pulmonary nodule † 0.4546 0.4475 0.4536 0.4328 0.4056 0.4498 0.4373 0.4711 0.5190 0.5082 Mediastinal widening 0.4516 0.2910 0.5210 0.5626 0.3015 0.3110 0.0671 0.5661 0.5434 0.6183 Improved aeration* 0.1794 0.1320 0.1788 0.2321 0.2294 0.1187 0.0812 0.1630 0.1655 0.2088 Chest tube 0.7093 0.6263 0.8010 0.7736 0.5184 0.6333 0.2387 0.7641 0.7843 0.8197 Malignant nodules* 0.5660 0.5460 0.8378 0.7739 0.5613 0.5259 0.3489 0.7722 0.6435 0.8341 Tracheostomy tube † 0.8188 0.5510 0.9208 0.9286 0.6431 0.7836 0.3820 0.9091 0.9351 0.9420 Cancer 0.4182 0.3881 0.5186 0.5423 0.3229 0.3744 0.1783 0.5266 0.5250 0.5782 Pacemaker † 0.8909 0.7716 0.8867 0.8915 0.7061 0.7814 0.0913 0.8990 0.8785 0.9141 Mediastinal drains † 0.8287 0.8000 0.8596 0.8549 0.6937 0.7929 0.3916 0.8604 0.8412 0.8767 Atrial dilation* 0.4234 0.4285 0.4832 0.4259 0.3993 0.2540 0.2122 0.4778 0.4079 0.5216 AICD † 0.8898 0.8525 0.9022 0.9200 0.8301 0.8422 0.1752 0.9235 0.9142 0.9363 Subclavian line* 0.3550 0.2866 0.4000 0.5500 0.1976 0.2930 0.1766 0.5772 0.6610 0.7662 Emphysema † 0.6585 0.5354 0.7088 0.7458 0.4848 0.6519 0.1697 0.7421 0.7256 0.7681 Interstitial edema 0.3874 0.3438 0.4495 0.4102 0.2809 0.3869 0.1943 0.4397 0.3883 0.4538 Hydropneumothorax* 0.6851 0.5126 0.7496 0.7507 0.3749 0.6762 0.2787 0.8153 0.7151 0.8101 Diaphragmatic hernia † 0.5304 0.4191 0.5711 0.6319 0.4860 0.4490 0.1960 0.6365 0.5939 0.6874 Pericardial effusion* 0.4782 0.3706 0.5595 0.5666 0.3566 0.3413 0.1219 0.6014 0.5062 0.5509 Healed rib fracture † 0.3714 0.3512 0.5636 0.5526 0.3376 0.3717 0.2702 0.5980 0.4132 0.6093 Alveolar edema † 0.5984 0.5491 0.6457 0.6229 0.4589 0.5946 0.3493 0.6652 0.5918 0.6406 Aortic valve replacement* 0.5872 0.6172 0.6680 0.6853 0.4824 0.5758 0.2655 0.6779 0.6181 0.6875 Hilar enlargement † 0.3339 0.2277 0.5051 0.5885 0.1850 0.2701 0.1949 0.5914 0.5469 0.6339 Internal jugular line 0.5418 0.4932 0.6264 0.6573 0.4222 0.5850 0.2929 0.6705 0.6502 0.6836 Normal 0.6923 0.6841 0.7199 0.7171 0.6225 0.6828 0.6225 0.7062 0.7128 0.7105 Diffused nodules † 0.5820 0.4851 0.8023 0.7611 0.5481 0.5681 0.2928 0.7464 0.7085 0.8431 Calcified nodules* 0.4547 0.3966 0.5516 0.4449 0.3933 0.4464 0.4102 0.4854 0.5287 0.6175 Clavicle fracture † 0.3627 0.3889 0.4623 0.4197 0.2760 0.3249 0.1934 0.7188 0.3951 0.4773 Nasogastric tube 0.7744 0.7678 0.7969 0.7987 0.6954 0.7845 0.6970 0.8060 0.7785 0.8286 Central venous line 0.4263 0.3848 0.5412 0.5475 0.3366 0.4237 0.1055 0.5228 0.5659 0.5785 Left subclavian line † 0.4680 0.3878 0.5512 0.6506 0.3149 0.4088 0.2338 0.5643 0.7287 0.7906 Pneumomediastinum* 0.5265 0.3350 0.5852 0.5787 0.4518 0.4560 0.0219 0.5172 0.5479 0.6627 Interstitial thickening † 0.3795 0.3608 0.4101 0.3733 0.2708 0.4132 0.0497 0.4081 0.3364 0.4018 Pulmonary vascular congestion 0.3164 0.2440 0.3666 0.3450 0.2304 0.2904 0.1417 0.3391 0.2984 0.3628 Swan-Ganz catheter 0.6314 0.6129 0.8221 0.7864 0.6967 0.6862 0.4623 0.8068 0.8209 0.8579 Aeration* 0.2068 0.2522 0.1746 0.1633 0.1129 0.1153 0.0363 0.1597 0.2466 0.2361 Mediastinal shift* 0.7120 0.5159 0.7338 0.8226 0.6289 0.5757 0.1289 0.8206 0.7639 0.8086 Surgical hardware* 0.2421 0.2686 0.3766 0.2718 0.1718 0.2746 0.0787 0.3165 0.4028 0.4883 Patchy opacities* 0.4467 0.3387 0.4376 0.4781 0.2820 0.4586 0.2962 0.5235 0.4881 0.5067 Mediastinal clips † 0.5951 0.5597 0.6651 0.6694 0.4902 0.5801 0.0343 0.6774 0.6418 0.6817 Opacity 0.3001 0.2787 0.3443 0.3332 0.2229 0.2803 0.2508 0.3253 0.3042 0.3328 Pulmonary hypertension* 0.3944 0.2911 0.4159 0.5965 0.2609 0.2544 0.0988 0.6268 0.5283 0.6062 Atelectasis 0.3001 0.2742 0.3553 0.3545 0.2196 0.3046 0.2128 0.3441 0.3143 0.3465 Hyperinflation † 0.3276 0.2993 0.3463 0.3777 0.2704 0.3071 0.2068 0.3384 0.3152 0.3643 ILD 0.3557 0.2968 0.4108 0.4027 0.2290 0.3506 0.0930 0.4000 0.3629 0.4217 Continued on next page 59 Abnormality EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BioViL-T Med-CLIP Med-AR-2B (ours) Med-AR-8B (ours) COPD † 0.4606 0.4493 0.5218 0.5281 0.4481 0.4458 0.2224 0.5370 0.4905 0.5550 Opacification* 0.6150 0.4993 0.6474 0.5188 0.5452 0.4164 0.2851 0.5743 0.4452 0.6347 Mediport* 0.5721 0.4911 0.7987 0.7365 0.4156 0.3077 0.1064 0.7970 0.8466 0.9598 Lung transplant* 0.6246 0.4893 0.7299 0.7598 0.4528 0.5600 0.0564 0.7254 0.7837 0.8031 Surgical clips 0.2983 0.2065 0.4871 0.4295 0.2495 0.1909 0.1235 0.4662 0.5461 0.6546 Low lung volumes 0.5147 0.4747 0.5261 0.5259 0.4175 0.4736 0.2819 0.5373 0.5036 0.5263 Aortic aneurysm † 0.3686 0.1661 0.4663 0.5399 0.2996 0.2094 0.1696 0.5696 0.5874 0.6616 Consolidation 0.3639 0.3507 0.4064 0.3944 0.3095 0.3652 0.2980 0.3883 0.3556 0.3888 Pneumonia 0.3409 0.2651 0.4288 0.4376 0.2099 0.3418 0.1099 0.4088 0.3811 0.4219 Free subdiaphragmatic air* 0.3347 0.3187 0.3645 0.5128 0.2524 0.2971 0.2507 0.6795 0.5527 0.7349 Fibrosis 0.5626 0.5214 0.6411 0.6390 0.4480 0.5372 0.3430 0.6500 0.6057 0.6680 Lung mass † 0.4917 0.3919 0.6327 0.6337 0.3634 0.4169 0.2362 0.6623 0.5989 0.6744 Endotracheal tube 0.8185 0.7776 0.8449 0.8518 0.7407 0.8108 0.7292 0.8493 0.8308 0.8653 Pulmonary arterial hypertension signs* 0.5151 0.4258 0.5067 0.6775 0.3297 0.3318 0.2061 0.7153 0.5421 0.6492 Calcification 0.4798 0.4134 0.5284 0.5662 0.3701 0.4453 0.2272 0.5775 0.5186 0.6145 60 S5 MIMIC-CXR: Per-Label ResultsThe evaluated MIMIC-CXR vocabulary contains 110 labels. The available per-label results below cover 108 labels, in the order AUROC, AUPRC, sensitivity,specificity, and Youden’s J ; results for the remaining two labels are not included in these listings. S5.1 AUROCTable S7: Per-label AUROC comparison across baseline, contrastive, and autoregressive models on the MIMIC-CXR dataset. Tail abnormalities are markedwith an asterisk (*), and medium-prevalence abnormalities are marked with a dagger ( † ). For each row, the best value is shown in bold and the second-best value is underlined. Abnormality EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BioViL-T Med-CLIP Med-AR-2B (ours) Med-AR-8B (ours) Diaphragm Elevation 0.8174 0.8314 0.8680 0.9072 0.8812 0.6922 0.5298 0.9076 0.9209 0.9251 Parenchymal Opacities † 0.8764 0.8839 0.8993 0.8904 0.8877 0.8674 0.7626 0.8869 0.8893 0.8975 Nodules 0.7210 0.7590 0.8355 0.8155 0.7608 0.7351 0.5837 0.7949 0.8038 0.8303 Pleural Effusion 0.9117 0.9215 0.9339 0.9327 0.9134 0.9150 0.8240 0.9287 0.9303 0.9321 Congestive Heart Failure* 0.8128 0.8403 0.8747 0.8410 0.8649 0.7886 0.6731 0.8774 0.8726 0.8725 Support Devices 0.9173 0.9198 0.9325 0.9317 0.9246 0.9069 0.8886 0.9303 0.9348 0.9371 PICC Line 0.7570 0.8739 0.9749 0.9723 0.9571 0.7100 0.6885 0.9389 0.9703 0.9754 Pleural Thickening 0.7481 0.7853 0.8410 0.8406 0.7738 0.7292 0.6119 0.8408 0.8433 0.8538 Soft Tissue Swelling † 0.8362 0.8617 0.9043 0.8990 0.8810 0.8010 0.5775 0.9027 0.9033 0.9126 Neoplastic Nodules* 0.8326 0.8873 0.9634 0.9488 0.8887 0.8230 0.4787 0.9304 0.9391 0.9594 Hemothorax* 0.8496 0.8937 0.9294 0.9041 0.9219 0.7409 0.6832 0.9149 0.9345 0.9302 CABG † 0.8889 0.9219 0.9157 0.9281 0.9243 0.8536 0.5751 0.9256 0.9306 0.9432 Scarring* 0.7322 0.7442 0.8268 0.8343 0.7480 0.7398 0.4652 0.8227 0.8395 0.8484 Aortic Tortuosity 0.8132 0.8346 0.8635 0.8786 0.8442 0.7932 0.6592 0.8772 0.8784 0.8900 Solitary Nodules † 0.7742 0.7805 0.8414 0.8179 0.7723 0.7890 0.7273 0.8049 0.8140 0.8137 Tracheal Deviation* 0.7055 0.7214 0.7384 0.6920 0.6806 0.6645 0.5510 0.7526 0.7389 0.7424 Scoliosis † 0.8195 0.8073 0.8273 0.9076 0.8473 0.7352 0.6240 0.8858 0.8829 0.9284 Pulmonary Edema 0.8873 0.8901 0.9081 0.9045 0.8909 0.8888 0.8165 0.9005 0.9042 0.9050 Rib Fractures 0.6943 0.7363 0.8838 0.8701 0.7591 0.6720 0.5705 0.8621 0.8606 0.8859 ARDS* 0.9305 0.9402 0.9476 0.9314 0.9372 0.9439 0.8475 0.9472 0.9426 0.9521 Aspiration † 0.7102 0.7391 0.7864 0.7832 0.7507 0.6887 0.6111 0.7804 0.7879 0.7896 Flattened Diaphragm † 0.9090 0.9140 0.9245 0.9218 0.9218 0.8978 0.7031 0.9251 0.9240 0.9351 Reticular Opacities † 0.8402 0.8696 0.9070 0.8881 0.8614 0.8391 0.5574 0.8993 0.9071 0.9105 Kyphosis* 0.8730 0.8812 0.8981 0.8940 0.8923 0.8444 0.4894 0.9004 0.9083 0.9114 Vascular Redistribution Cephalization 0.7786 0.7890 0.8070 0.8020 0.7914 0.7805 0.6883 0.7966 0.7987 0.8058 Degenerative Changes † 0.7789 0.7810 0.7948 0.8149 0.7993 0.7531 0.5496 0.8097 0.8140 0.8232 Dialysis Catheter* 0.7905 0.9112 0.9534 0.9569 0.9218 0.6397 0.4393 0.9420 0.9600 0.9790 Pigtail Catheter* 0.7977 0.8362 0.9631 0.9432 0.9358 0.7714 0.6040 0.9423 0.9584 0.9721 Hilar Lymphadenopathy † 0.7063 0.7685 0.8482 0.8422 0.8114 0.6923 0.5031 0.8284 0.8314 0.8685 Chest Tube Removal* 0.7996 0.8497 0.8972 0.9000 0.8505 0.6606 0.5530 0.9063 0.9254 0.9248 Retrocardiac Opacity † 0.7026 0.7130 0.7747 0.7687 0.7471 0.6648 0.6476 0.7527 0.7441 0.7942 Pneumothorax 0.8648 0.9024 0.9651 0.9474 0.9175 0.8532 0.7191 0.9411 0.9452 0.9555 Vascular Congestion † 0.7411 0.7589 0.7914 0.7641 0.7649 0.7543 0.7114 0.7642 0.7755 0.7882 Esophageal Drainage Tube* 0.8963 0.9067 0.9157 0.9070 0.8919 0.8905 0.7115 0.9016 0.9119 0.9153 Perihilar Opacities 0.7824 0.8014 0.8393 0.8293 0.8201 0.7870 0.6902 0.8259 0.8380 0.8488 Interstitial Lung Disease Pattern 0.7676 0.7848 0.8281 0.8271 0.7907 0.7685 0.6256 0.8205 0.8236 0.8283 Extubation* 0.7936 0.8347 0.8273 0.7982 0.7890 0.7898 0.6817 0.8376 0.8847 0.8790 Sternotomy Wires 0.9589 0.9626 0.9654 0.9650 0.9623 0.9509 0.6693 0.9644 0.9624 0.9662 Continued on next page 61 Abnormality EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BioViL-T Med-CLIP Med-AR-2B (ours) Med-AR-8B (ours) MPA Enlargement † 0.7038 0.7326 0.7805 0.8474 0.7650 0.6995 0.6166 0.8237 0.8172 0.8558 Subcutaneous Emphysema † 0.8926 0.9405 0.9771 0.9731 0.9483 0.8331 0.6448 0.9599 0.9674 0.9803 Bronchial Wall Thickening* 0.7431 0.7400 0.7800 0.7756 0.7155 0.6695 0.5941 0.7721 0.7727 0.7826 Chest Wall Deformity † 0.7617 0.7697 0.8540 0.8560 0.7748 0.7093 0.5840 0.8630 0.8694 0.8858 Feeding Tube 0.9079 0.9134 0.9277 0.9316 0.9333 0.9138 0.8944 0.9241 0.9374 0.9428 Fluid Overload* 0.7961 0.7822 0.8050 0.7746 0.7824 0.7516 0.7185 0.8065 0.8100 0.8104 Sternotomy † 0.9556 0.9605 0.9643 0.9636 0.9607 0.9465 0.5883 0.9620 0.9638 0.9671 Granuloma † 0.7294 0.7150 0.8355 0.7975 0.7054 0.6857 0.6099 0.8273 0.8457 0.8857 Mediastinal Mass* 0.6958 0.7525 0.8352 0.8261 0.7662 0.6407 0.4884 0.8406 0.8385 0.8699 Artifacts 0.7019 0.6963 0.7347 0.7380 0.7255 0.6549 0.6201 0.7287 0.7492 0.7526 Spinal Hardware* 0.6375 0.6018 0.9361 0.8704 0.7627 0.5862 0.5685 0.8257 0.8979 0.9415 Cardiomegaly 0.8444 0.8376 0.8705 0.8686 0.8556 0.8215 0.7114 0.8667 0.8699 0.8716 Solitary Pulmonary Nodule † 0.7780 0.7875 0.8235 0.8103 0.7629 0.7759 0.6806 0.7919 0.8019 0.8182 Empyema* 0.8684 0.9025 0.9247 0.9420 0.8869 0.7805 0.7416 0.9488 0.9504 0.9546 Mediastinal Widening 0.8017 0.7990 0.8691 0.8617 0.8299 0.7572 0.7107 0.8698 0.8750 0.8836 Chest Tube 0.8918 0.9456 0.9748 0.9659 0.9535 0.8629 0.7221 0.9590 0.9715 0.9766 Malignant Nodules* 0.7996 0.8742 0.9547 0.9270 0.8725 0.8042 0.6165 0.9170 0.9267 0.9428 Pacemaker 0.9696 0.9773 0.9801 0.9785 0.9796 0.8988 0.7083 0.9808 0.9822 0.9818 Cancer 0.8119 0.8436 0.8830 0.8844 0.8538 0.7775 0.6163 0.8736 0.8865 0.8955 Tracheostomy Tube † 0.8942 0.9475 0.9847 0.9890 0.9720 0.8169 0.7736 0.9873 0.9903 0.9921 Atrial Dilation* 0.7413 0.7190 0.7609 0.7484 0.7221 0.5786 0.4470 0.7372 0.7825 0.7747 MPA Dilation † 0.7178 0.7528 0.8000 0.8679 0.7911 0.7220 0.5724 0.8594 0.8455 0.8823 Emphysema † 0.8826 0.8989 0.9150 0.9262 0.9007 0.8770 0.7397 0.9223 0.9218 0.9284 Interstitial Edema 0.8334 0.8371 0.8670 0.8586 0.8427 0.8414 0.7330 0.8573 0.8613 0.8596 Diaphragmatic Hernia † 0.8046 0.8076 0.8833 0.9248 0.8429 0.7396 0.5169 0.8662 0.9022 0.9296 Cavitary Nodule* 0.8067 0.8881 0.9008 0.8959 0.8426 0.8190 0.6956 0.8906 0.8873 0.9082 Pericardial Effusion † 0.8215 0.8216 0.8946 0.8896 0.8681 0.7792 0.6444 0.8936 0.8903 0.9034 Healed Rib Fracture † 0.7173 0.7517 0.9171 0.8931 0.7664 0.7299 0.5865 0.9032 0.8752 0.9166 Alveolar Edema † 0.8824 0.8997 0.9104 0.9041 0.9011 0.8991 0.8039 0.9149 0.9040 0.9054 Aortic Valve Replacement* 0.9356 0.9390 0.9566 0.9444 0.9206 0.9403 0.5173 0.9499 0.9484 0.9630 Hilar Enlargement 0.7304 0.7735 0.8305 0.8382 0.8038 0.7395 0.5691 0.8446 0.8482 0.8522 Hiatal Hernia* 0.8251 0.8795 0.8851 0.9465 0.8751 0.8033 0.6499 0.8924 0.9470 0.9702 Bronchovascular Crowding* 0.8453 0.8708 0.8907 0.8694 0.8696 0.8333 0.6384 0.8885 0.8835 0.8847 Internal Jugular Line † 0.8619 0.8858 0.9409 0.9546 0.9515 0.8651 0.8201 0.9497 0.9612 0.9619 Normal 0.9152 0.9240 0.9327 0.9341 0.9191 0.9137 0.8795 0.9306 0.9322 0.9328 Diffused Nodules* 0.8137 0.8516 0.9426 0.9166 0.8750 0.8460 0.6252 0.9123 0.9159 0.9382 Calcified Nodules* 0.6006 0.6488 0.8096 0.8045 0.7117 0.5008 0.5407 0.8446 0.8663 0.8518 Bibasilar Opacities † 0.7493 0.7758 0.8111 0.7916 0.7710 0.7158 0.6394 0.8157 0.8042 0.8203 Clavicle Fracture* 0.6127 0.6351 0.7335 0.7284 0.7303 0.6418 0.5694 0.8176 0.7346 0.7491 Nasogastric Tube 0.9307 0.9369 0.9477 0.9441 0.9412 0.9324 0.9150 0.9391 0.9530 0.9560 Central Venous Line † 0.7702 0.8704 0.9109 0.9152 0.9055 0.7396 0.7029 0.9076 0.9257 0.9310 Left Subclavian Line* 0.8428 0.8999 0.9335 0.9340 0.9345 0.8406 0.5907 0.8950 0.9591 0.9783 Pneumomediastinum* 0.8064 0.8175 0.9076 0.9040 0.8441 0.8014 0.4177 0.8597 0.8838 0.8930 Interstitial Thickening † 0.8452 0.8491 0.8810 0.8756 0.8459 0.8386 0.6379 0.8733 0.8752 0.8749 Dobhoff Tube † 0.9329 0.9616 0.9775 0.9720 0.9704 0.9491 0.8709 0.9598 0.9825 0.9848 Bronchiectasis* 0.8817 0.9004 0.9100 0.9023 0.8668 0.8798 0.6978 0.8736 0.8991 0.9219 Swan Ganz Catheter † 0.9378 0.9607 0.9873 0.9750 0.9728 0.9073 0.7127 0.9778 0.9888 0.9915 Bullae* 0.7336 0.8355 0.8572 0.8731 0.8366 0.7760 0.4163 0.8672 0.8606 0.9181 Adenocarcinoma* 0.7173 0.7907 0.8307 0.8171 0.8139 0.6909 0.4593 0.8549 0.8571 0.8617 Mediastinal Shift † 0.8841 0.8904 0.9202 0.9335 0.9202 0.8194 0.6301 0.9402 0.9438 0.9525 Mediastinal Clips* 0.9599 0.9446 0.9669 0.9676 0.9642 0.9415 0.4985 0.9663 0.9679 0.9686 Opacity 0.7268 0.7483 0.7739 0.7656 0.7504 0.7357 0.6760 0.7576 0.7621 0.7718 Atelectasis 0.7907 0.8041 0.8373 0.8278 0.8065 0.7927 0.6897 0.8231 0.8296 0.8339 Hyperinflation 0.8987 0.9023 0.9082 0.9130 0.9020 0.8922 0.7943 0.9122 0.9135 0.9171 Continued on next page 62 Abnormality EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BioViL-T Med-CLIP Med-AR-2B (ours) Med-AR-8B (ours) Diffuse Fibrosis* 0.9654 0.9712 0.9875 0.9879 0.9897 0.9672 0.6987 0.9929 0.9896 0.9921 Surgical Clips 0.6927 0.6827 0.8467 0.7927 0.8089 0.6706 0.6009 0.7727 0.8765 0.9166 Low Lung Volumes 0.8456 0.8471 0.8610 0.8646 0.8619 0.8211 0.7091 0.8619 0.8644 0.8676 Aortic Aneurysm* 0.5884 0.7299 0.8056 0.8483 0.7688 0.5658 0.4683 0.8076 0.8244 0.8460 Pneumonia 0.7498 0.7913 0.8314 0.8238 0.7945 0.7725 0.6492 0.8156 0.8232 0.8289 Consolidation 0.7789 0.8102 0.8457 0.8376 0.8106 0.7913 0.7064 0.8295 0.8330 0.8382 Free Subdiaphragmatic Air* 0.7672 0.8137 0.8534 0.8719 0.8642 0.7187 0.5412 0.8861 0.9005 0.9334 Thoracic Spine Degenerative Changes* 0.8280 0.8035 0.8342 0.8228 0.7926 0.7868 0.6204 0.8178 0.8364 0.8470 Port-a-Cath † 0.7986 0.9288 0.9827 0.9778 0.9758 0.6577 0.4950 0.9790 0.9871 0.9885 Fibrosis 0.7999 0.8371 0.8888 0.8842 0.8345 0.7953 0.7067 0.8792 0.8847 0.8919 Lung Mass † 0.7902 0.8710 0.9127 0.9127 0.8740 0.7576 0.5593 0.9041 0.9092 0.9149 Endotracheal Tube 0.9637 0.9699 0.9751 0.9750 0.9730 0.9664 0.9499 0.9725 0.9781 0.9794 Pulmonary Arterial Hypertension Signs † 0.7149 0.7699 0.7881 0.8282 0.7778 0.7361 0.6329 0.8373 0.8352 0.8538 Effusion 0.9118 0.9213 0.9337 0.9326 0.9127 0.9142 0.8221 0.9282 0.9301 0.9322 Calcification 0.7826 0.7938 0.8460 0.8668 0.8142 0.7568 0.6730 0.8668 0.8687 0.8835 Degenerative Spine Changes † 0.8031 0.8204 0.8447 0.8525 0.8355 0.7811 0.5151 0.8405 0.8649 0.8637 63 S5.2 AUPRCTable S8: Per-label AUPRC comparison across baseline, contrastive, and autoregressive models on the MIMIC-CXR dataset. Tail abnormalities are markedwith an asterisk (*), and medium-prevalence abnormalities are marked with a dagger ( † ). For each row, the best value is shown in bold and the second-best value is underlined. Abnormality EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BioViL-T Med-CLIP Med-AR-2B (ours) Med-AR-8B (ours) Diaphragm Elevation 0.2024 0.2282 0.3259 0.4529 0.3923 0.0722 0.0334 0.4595 0.4844 0.4967 Parenchymal Opacities † 0.0581 0.0552 0.0691 0.0632 0.0583 0.0502 0.0192 0.0549 0.0630 0.0633 Nodules 0.0991 0.1335 0.3179 0.2356 0.1577 0.1118 0.0422 0.2503 0.2624 0.2965 Pleural Effusion 0.8302 0.8463 0.8706 0.8682 0.8332 0.8343 0.6555 0.8605 0.8663 0.8674 Congestive Heart Failure* 0.0179 0.0190 0.0242 0.0186 0.0197 0.0113 0.0049 0.0241 0.0210 0.0220 Support Devices 0.1770 0.1640 0.1957 0.2010 0.1856 0.1477 0.1401 0.1972 0.1978 0.2053 Picc Line 0.1270 0.2494 0.6183 0.6005 0.5522 0.0928 0.0862 0.4965 0.6083 0.6171 Pleural Thickening 0.0715 0.0999 0.1886 0.2010 0.1159 0.0725 0.0373 0.2072 0.2294 0.2429 Soft Tissue Swelling † 0.2236 0.3382 0.4742 0.4699 0.4079 0.2401 0.0202 0.4624 0.4918 0.5041 Neoplastic Nodules* 0.0440 0.0737 0.4225 0.3393 0.1939 0.0999 0.0040 0.3755 0.3528 0.4110 Hemothorax* 0.0359 0.0357 0.0906 0.0704 0.0825 0.0086 0.0065 0.0838 0.1106 0.1075 CABG † 0.0965 0.1074 0.1105 0.1263 0.1185 0.0882 0.0143 0.1271 0.1446 0.1604 Scarring* 0.0080 0.0134 0.0183 0.0242 0.0095 0.0081 0.0028 0.0241 0.0209 0.0247 Aortic Tortuosity 0.1731 0.2006 0.2460 0.2945 0.2301 0.1600 0.0744 0.2798 0.2807 0.2965 Solitary Nodules † 0.0177 0.0200 0.0510 0.0407 0.0222 0.0212 0.0120 0.0442 0.0463 0.0438 Tracheal Deviation* 0.0070 0.0086 0.0260 0.0086 0.0094 0.0069 0.0040 0.0239 0.0228 0.0229 Scoliosis † 0.1075 0.1002 0.1150 0.4239 0.1843 0.0362 0.0198 0.3310 0.3032 0.4370 Pulmonary Edema 0.6478 0.6586 0.7119 0.7053 0.6661 0.6458 0.4713 0.6905 0.7007 0.7026 Rib Fractures 0.0579 0.0754 0.3188 0.2800 0.1034 0.0501 0.0310 0.2826 0.2635 0.3110 ARDS* 0.0800 0.0736 0.0882 0.0973 0.1259 0.0759 0.0328 0.1107 0.0987 0.1148 Aspiration † 0.0426 0.0467 0.0718 0.0671 0.0584 0.0329 0.0239 0.0645 0.0814 0.0785 Flattened Diaphragm † 0.0825 0.1012 0.1015 0.1308 0.1165 0.0765 0.0117 0.1266 0.1236 0.1373 Reticular Opacities † 0.0698 0.0788 0.1119 0.1400 0.1176 0.0699 0.0065 0.1208 0.1262 0.1446 Kyphosis* 0.0735 0.0511 0.0932 0.1243 0.1253 0.0323 0.0036 0.1194 0.1249 0.1538 Vascular Redistribution Cephalization 0.1394 0.1474 0.1608 0.1561 0.1537 0.1414 0.0858 0.1522 0.1575 0.1555 Degenerative Changes † 0.0399 0.0391 0.0436 0.0541 0.0469 0.0329 0.0141 0.0443 0.0486 0.0539 Dialysis Catheter* 0.0113 0.0666 0.2079 0.2458 0.2383 0.0057 0.0031 0.1679 0.2671 0.3351 Pigtail Catheter* 0.0257 0.0351 0.2607 0.1674 0.2067 0.0234 0.0101 0.1878 0.2788 0.2568 Hilar Lymphadenopathy † 0.0244 0.0383 0.1049 0.0985 0.0791 0.0225 0.0089 0.0941 0.0933 0.1192 Chest Tube Removal* 0.0143 0.0242 0.0390 0.0669 0.0545 0.0060 0.0038 0.1120 0.1385 0.1353 Retrocardiac Opacity † 0.0283 0.0189 0.0369 0.0371 0.0248 0.0136 0.0133 0.0272 0.0267 0.0351 Pneumothorax 0.2790 0.3834 0.7375 0.5616 0.4513 0.2739 0.0993 0.5777 0.5707 0.6830 Vascular Congestion † 0.0243 0.0264 0.0286 0.0261 0.0331 0.0242 0.0206 0.0261 0.0381 0.0280 Esophageal Drainage Tube* 0.0188 0.0189 0.0249 0.0199 0.0210 0.0227 0.0085 0.0267 0.0211 0.0231 Perihilar Opacities 0.0786 0.0867 0.1473 0.1374 0.1277 0.0741 0.0458 0.1404 0.1621 0.1707 Interstitial Lung Disease Pattern 0.3166 0.3426 0.4286 0.4287 0.3772 0.3266 0.1487 0.4176 0.4229 0.4301 Extubation* 0.0153 0.0178 0.0145 0.0191 0.0118 0.0114 0.0066 0.0158 0.0264 0.0267 Sternotomy Wires 0.3341 0.3315 0.3415 0.3462 0.3298 0.2880 0.0494 0.3465 0.3259 0.3451 MPA Enlargement † 0.0201 0.0229 0.0535 0.1105 0.0471 0.0178 0.0104 0.0972 0.1020 0.1529 Subcutaneous Emphysema † 0.1084 0.1976 0.2747 0.3390 0.2398 0.1028 0.0137 0.2627 0.3296 0.3153 Bronchial Wall Thickening* 0.0160 0.0176 0.0234 0.0260 0.0235 0.0108 0.0066 0.0231 0.0240 0.0309 Chest Wall Deformity † 0.0347 0.0357 0.0665 0.0871 0.0436 0.0281 0.0107 0.0809 0.0780 0.1031 Feeding Tube 0.1501 0.1602 0.2075 0.2156 0.2364 0.1604 0.1349 0.2123 0.2573 0.2866 Fluid Overload* 0.0187 0.0171 0.0163 0.0156 0.0146 0.0108 0.0099 0.0170 0.0187 0.0182 Sternotomy † 0.2050 0.2170 0.2365 0.2266 0.2191 0.1922 0.0204 0.2429 0.2342 0.2571 Granuloma † 0.0294 0.0223 0.1820 0.1298 0.0270 0.0200 0.0125 0.1870 0.2335 0.2724 Continued on next page 64 Abnormality EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BioViL-T Med-CLIP Med-AR-2B (ours) Med-AR-8B (ours) Mediastinal Mass* 0.0126 0.0199 0.0998 0.1073 0.0596 0.0093 0.0055 0.1212 0.1190 0.1487 Artifacts 0.1160 0.0972 0.1334 0.1487 0.1304 0.0800 0.0665 0.1341 0.1497 0.1476 Spinal Hardware* 0.0085 0.0039 0.1289 0.0907 0.0464 0.0039 0.0032 0.0709 0.1165 0.1727 Cardiomegaly 0.6932 0.6877 0.7447 0.7424 0.7206 0.6526 0.4592 0.7397 0.7453 0.7489 Solitary Pulmonary Nodule † 0.0213 0.0241 0.0502 0.0456 0.0250 0.0240 0.0121 0.0394 0.0386 0.0558 Empyema* 0.0173 0.0307 0.0562 0.0581 0.0510 0.0085 0.0072 0.0654 0.0823 0.0801 Mediastinal Widening 0.1222 0.1181 0.1975 0.1884 0.1438 0.0754 0.0564 0.2021 0.2236 0.2368 Chest Tube 0.2596 0.4147 0.5607 0.4717 0.4508 0.2058 0.0555 0.4424 0.5292 0.5425 Malignant Nodules* 0.0440 0.0863 0.3912 0.2806 0.1655 0.0418 0.0077 0.3163 0.3028 0.3769 Pacemaker 0.3571 0.3789 0.4491 0.4153 0.4511 0.1913 0.0498 0.4559 0.4756 0.4842 Cancer 0.1648 0.2177 0.3205 0.3023 0.2554 0.1267 0.0485 0.2873 0.3141 0.3464 Tracheostomy Tube † 0.2039 0.3889 0.7064 0.7306 0.6461 0.0795 0.0606 0.6866 0.7437 0.7593 Atrial Dilation* 0.0071 0.0062 0.0118 0.0091 0.0125 0.0044 0.0023 0.0106 0.0154 0.0141 MPA Dilation † 0.0169 0.0192 0.0412 0.0891 0.0271 0.0139 0.0068 0.0851 0.0911 0.1704 Emphysema † 0.2063 0.2257 0.2655 0.2822 0.2404 0.1691 0.0579 0.2716 0.2903 0.3274 Interstitial Edema 0.1984 0.2047 0.2634 0.2527 0.2311 0.2185 0.1012 0.2515 0.2562 0.2626 Diaphragmatic Hernia † 0.0743 0.0876 0.2256 0.4658 0.1330 0.0457 0.0143 0.1711 0.3807 0.5783 Cavitary Nodule* 0.0169 0.0248 0.0367 0.0298 0.0276 0.0197 0.0063 0.0382 0.0461 0.0483 Pericardial Effusion † 0.0855 0.0459 0.1639 0.1606 0.1104 0.0250 0.0120 0.1662 0.1823 0.1926 Healed Rib Fracture † 0.0440 0.0539 0.3182 0.2653 0.0635 0.0437 0.0195 0.2968 0.2553 0.3444 Alveolar Edema † 0.0540 0.0772 0.0942 0.0774 0.0826 0.0785 0.0215 0.0904 0.0952 0.1089 Aortic Valve Replacement* 0.0428 0.0396 0.0799 0.0405 0.0428 0.0423 0.0039 0.0546 0.0500 0.0882 Hilar Enlargement 0.0709 0.0880 0.1893 0.2125 0.1402 0.0686 0.0283 0.2023 0.2211 0.2277 Hiatal Hernia* 0.0142 0.0204 0.0685 0.1759 0.0607 0.0108 0.0041 0.0413 0.1997 0.2812 Bronchovascular Crowding* 0.0305 0.0406 0.0581 0.0475 0.0525 0.0334 0.0095 0.0390 0.0452 0.0465 Internal Jugular Line † 0.0811 0.0907 0.1694 0.2193 0.2237 0.0909 0.0765 0.2034 0.2518 0.2449 Normal 0.7137 0.7445 0.7720 0.7757 0.7237 0.7137 0.6128 0.7662 0.7678 0.7723 Diffused Nodules* 0.0437 0.0896 0.4378 0.2301 0.1839 0.0996 0.0082 0.3240 0.3656 0.4089 Calcified Nodules* 0.0037 0.0043 0.0559 0.0500 0.0125 0.0025 0.0028 0.1021 0.1101 0.1176 Bibasilar Opacities † 0.0190 0.0211 0.0385 0.0274 0.0291 0.0149 0.0103 0.0492 0.0319 0.0374 Clavicle Fracture* 0.0077 0.0101 0.0187 0.0217 0.0246 0.0092 0.0068 0.1386 0.0198 0.0441 Nasogastric Tube 0.3462 0.3706 0.4428 0.4220 0.4190 0.3627 0.3023 0.3778 0.4899 0.5046 Central Venous Line † 0.0490 0.0926 0.1664 0.1507 0.1576 0.0478 0.0464 0.1627 0.1944 0.2154 Left Subclavian Line* 0.0157 0.0361 0.0432 0.0862 0.0987 0.0137 0.0048 0.0293 0.1395 0.1675 Pneumomediastinum* 0.0656 0.0696 0.1993 0.1967 0.1439 0.0896 0.0038 0.1281 0.1691 0.2008 Interstitial Thickening † 0.1221 0.1276 0.1771 0.1609 0.1563 0.1367 0.0238 0.1646 0.1648 0.1624 Dobhoff Tube † 0.1245 0.1956 0.3380 0.2584 0.2561 0.1279 0.0359 0.1803 0.3751 0.3676 Bronchiectasis* 0.0413 0.0586 0.0914 0.0659 0.0492 0.0332 0.0061 0.0681 0.1246 0.1049 Swan Ganz Catheter † 0.1580 0.2207 0.5597 0.5546 0.6003 0.1058 0.0228 0.5337 0.6929 0.7261 Bullae* 0.0123 0.0247 0.0360 0.0588 0.0183 0.0091 0.0023 0.0449 0.0353 0.0829 Adenocarcinoma* 0.0085 0.0102 0.0102 0.0106 0.0089 0.0055 0.0024 0.0132 0.0184 0.0169 Mediastinal Shift † 0.1671 0.0893 0.2091 0.2400 0.2283 0.0352 0.0137 0.2888 0.2692 0.3050 Mediastinal Clips* 0.1378 0.0947 0.1650 0.1451 0.1501 0.0967 0.0051 0.1405 0.1368 0.1757 Opacity 0.1334 0.1479 0.1751 0.1707 0.1604 0.1352 0.1049 0.1650 0.1793 0.1825 Atelectasis 0.6414 0.6578 0.7243 0.7034 0.6697 0.6483 0.4755 0.6993 0.7106 0.7159 Hyperinflation 0.3959 0.4022 0.4282 0.4505 0.4165 0.3645 0.1743 0.4304 0.4501 0.4628 Diffuse Fibrosis* 0.2105 0.2133 0.3875 0.3840 0.3293 0.1707 0.0050 0.3917 0.3602 0.4031 Surgical Clips 0.0469 0.0436 0.2229 0.1270 0.1819 0.0409 0.0269 0.0944 0.2920 0.3603 Low Lung Volumes 0.4624 0.4720 0.5111 0.5258 0.5170 0.4089 0.2274 0.5136 0.5208 0.5305 Aortic Aneurysm* 0.0077 0.0234 0.1010 0.1217 0.0778 0.0064 0.0043 0.1132 0.1589 0.1846 Pneumonia 0.3045 0.3679 0.4704 0.4476 0.3944 0.3430 0.2012 0.4396 0.4585 0.4695 Consolidation 0.3330 0.3765 0.4532 0.4340 0.3944 0.3471 0.2474 0.4229 0.4367 0.4469 Free Subdiaphragmatic Air* 0.0181 0.0293 0.1464 0.1476 0.0770 0.0118 0.0052 0.2848 0.3195 0.5151 Thoracic Spine Degenerative Changes* 0.0193 0.0199 0.0260 0.0274 0.0183 0.0165 0.0062 0.0276 0.0478 0.0434 Continued on next page 65 Abnormality EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BioViL-T Med-CLIP Med-AR-2B (ours) Med-AR-8B (ours) Port-a-cath † 0.1292 0.2633 0.4629 0.4219 0.4336 0.0299 0.0155 0.4716 0.5031 0.5291 Fibrosis 0.1511 0.2001 0.3441 0.3357 0.2472 0.1443 0.0918 0.3127 0.3278 0.3478 Lung Mass † 0.0726 0.1562 0.3144 0.2617 0.2027 0.0519 0.0195 0.2698 0.3075 0.3058 Endotracheal Tube 0.6125 0.6637 0.7116 0.7063 0.7063 0.6301 0.5439 0.6814 0.7463 0.7660 Pulmonary Arterial Hypertension Signs † 0.0273 0.0412 0.0757 0.1364 0.0614 0.0263 0.0174 0.1348 0.1187 0.1734 Effusion 0.8306 0.8452 0.8699 0.8674 0.8315 0.8326 0.6480 0.8589 0.8648 0.8668 Calcification 0.2096 0.2263 0.3268 0.3681 0.2888 0.1772 0.1139 0.3885 0.4000 0.4273 Degenerative Spine Changes † 0.0270 0.0281 0.0313 0.0362 0.0294 0.0179 0.0071 0.0313 0.0366 0.0383 66 S5.3 SensitivityTable S9: Per-label Sensitivity comparison across baseline, contrastive, and autoregressive models on the MIMIC dataset. Tail abnormalities are marked with an asterisk (*), and medium-prevalence abnormalities are marked with a dagger ( † ). For each row, the best value is shown in bold and the second-best value is underlined. Abnormality EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BioViL-T Med-CLIP Med-AR-2B (ours) Med-AR-8B (ours) Diaphragm elevation 0.7113 0.7307 0.7942 0.8108 0.7555 0.7320 0.9171 0.8177 0.7887 0.8522 Parenchymal opacities † 0.8623 0.8922 0.8802 0.9222 0.8922 0.8144 0.7605 0.8443 0.8802 0.8323 Nodules 0.7195 0.6697 0.6868 0.7274 0.7235 0.6776 0.5531 0.6239 0.6645 0.7471 Pleural effusion 0.8516 0.8709 0.8747 0.8538 0.8735 0.8706 0.7845 0.8671 0.8535 0.8665 Congestive heart failure* 0.8095 0.8254 0.8571 0.8730 0.7619 0.9206 0.9206 0.8889 0.8413 0.8571 Support devices 0.9427 0.9301 0.9516 0.9498 0.9301 0.9409 0.9480 0.9570 0.9642 0.9749 PICC line 0.7813 0.8960 0.9595 0.9538 0.9143 0.7861 0.7813 0.8796 0.9566 0.9624 Pleural thickening 0.6724 0.7966 0.7483 0.7414 0.6776 0.6138 0.7845 0.7914 0.7879 0.8000 Soft tissue swelling † 0.6955 0.6925 0.7910 0.7522 0.8000 0.6537 0.8448 0.7612 0.8209 0.8149 Neoplastic nodules* 0.7604 0.7917 0.9271 0.8646 0.7292 0.7188 0.9792 0.9375 0.8750 0.8854 Hemothorax* 0.8140 0.8837 0.9302 0.8140 0.8721 0.8372 0.8721 0.8605 0.9419 0.8605 CABG † 0.7444 0.9000 0.9074 0.8889 0.8519 0.7148 0.7370 0.9111 0.9333 0.9370 Scarring* 0.8841 0.7246 0.8261 0.7536 0.7826 0.8406 0.7826 0.7971 0.9130 0.7681 Aortic tortuosity 0.7877 0.7777 0.7988 0.8290 0.7998 0.8390 0.5493 0.7988 0.8290 0.8390 Solitary nodules † 0.7518 0.8540 0.7591 0.7372 0.6788 0.9197 0.8540 0.7153 0.7664 0.6569 Tracheal deviation* 0.7927 0.8171 0.8049 0.6463 0.5610 0.9512 0.9146 0.8293 0.7561 0.7805 Scoliosis † 0.7204 0.8059 0.7303 0.7730 0.7007 0.7533 0.5855 0.7467 0.7632 0.8717 Pulmonary edema 0.8645 0.8352 0.8822 0.8778 0.8341 0.8290 0.8157 0.8724 0.8664 0.8854 Rib fractures 0.6524 0.6747 0.7945 0.8459 0.5908 0.6815 0.8339 0.7842 0.7380 0.7568 ARDS* 0.9176 0.9176 0.9176 0.8353 0.8824 0.9294 0.8471 0.9529 0.9059 0.8588 Aspiration † 0.7408 0.7824 0.8264 0.7359 0.7090 0.8166 0.8655 0.7408 0.7482 0.7800 Flattened diaphragm † 0.8917 0.9045 0.9108 0.8599 0.8471 0.8726 0.8471 0.9299 0.8217 0.8917 Reticular opacities † 0.7939 0.8702 0.8397 0.7786 0.7634 0.7405 0.8092 0.8855 0.9237 0.8321 Kyphosis* 0.7590 0.8916 0.9518 0.7590 0.8675 0.8193 0.7711 0.8313 0.8434 0.7711 Vascular redistribution cephalization 0.8651 0.7933 0.8083 0.8225 0.8554 0.7941 0.8394 0.8705 0.8571 0.8403 Degenerative changes † 0.8526 0.7649 0.8140 0.8105 0.8316 0.7544 0.7018 0.8211 0.8526 0.8667 Dialysis catheter* 0.7037 0.8519 0.9259 0.8765 0.8395 0.7654 0.9877 0.9136 0.9259 0.9630 Pigtail catheter* 0.8642 0.8765 0.9198 0.9259 0.8642 0.8086 0.7469 0.8889 0.8951 0.9321 Hilar lymphadenopathy † 0.7241 0.6207 0.8276 0.7438 0.7340 0.7438 0.7931 0.8128 0.7537 0.7685 Chest tube removal* 0.7671 0.9041 0.8219 0.9041 0.7397 0.9589 0.5205 0.8904 0.8904 0.7808 Retrocardiac opacity † 0.7374 0.7765 0.7039 0.6872 0.9050 0.6872 0.7933 0.7207 0.7263 0.7095 Pneumothorax 0.8103 0.8670 0.8876 0.8639 0.8557 0.8206 0.6649 0.8784 0.8753 0.8845 Vascular congestion † 0.7682 0.8545 0.8636 0.8000 0.7682 0.8136 0.7318 0.8409 0.9273 0.8455 Esophageal drainage tube* 0.9620 1.0000 0.9747 0.9873 0.9367 0.9747 0.7089 0.9620 0.9747 1.0000 Perihilar opacities 0.7681 0.7504 0.8319 0.8053 0.7965 0.7504 0.7327 0.7133 0.7416 0.7894 Interstitial lung disease pattern 0.7631 0.7649 0.7592 0.7584 0.6228 0.6854 0.7462 0.7336 0.7718 0.7475 Extubation* 0.9863 0.8356 0.8904 0.7123 0.8082 0.8082 0.7808 0.8630 0.9452 0.8356 Sternotomy wires 0.9647 0.9775 0.9952 0.9952 0.9759 0.9599 0.7095 0.9888 0.9936 0.9968 MPA enlargement † 0.8323 0.7081 0.7826 0.7702 0.6708 0.5901 0.7950 0.7081 0.6584 0.7081 Subcutaneous emphysema † 0.7826 0.8509 0.9379 0.9379 0.8696 0.7329 0.8323 0.9130 0.9006 0.9379 Bronchial wall thickening* 0.6857 0.6190 0.7238 0.6762 0.7714 0.7333 0.6286 0.7048 0.7905 0.7429 Chest wall deformity † 0.5939 0.7259 0.7360 0.8071 0.6599 0.7462 0.7919 0.8629 0.8274 0.8376 Feeding tube 0.9699 0.9826 0.9873 0.9953 0.9858 0.9731 0.9177 0.9905 0.9889 0.9937 Fluid overload* 0.7522 0.8584 0.9381 0.8230 0.8230 0.8850 0.8142 0.9381 0.8053 0.7434 Sternotomy † 0.9501 0.9948 0.9895 0.9948 0.9869 0.9423 0.8451 0.9869 0.9869 0.9895 Continued on next page 67 Abnormality EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BioViL-T Med-CLIP Med-AR-2B (ours) Med-AR-8B (ours) Granuloma † 0.8019 0.6415 0.6509 0.7783 0.6179 0.5849 0.7689 0.7264 0.7594 0.7830 Mediastinal mass* 0.8450 0.7132 0.7519 0.6512 0.6589 0.7519 0.8605 0.6899 0.6977 0.7829 Artifacts 0.7089 0.6509 0.7640 0.7139 0.7404 0.6942 0.6785 0.7325 0.6696 0.6775 Spinal hardware* 0.4098 0.5902 0.8525 0.6721 0.6557 0.3279 0.9180 0.6885 0.7377 0.7869 Cardiomegaly 0.8108 0.8111 0.8124 0.7796 0.7787 0.8048 0.8306 0.7998 0.8338 0.7971 Solitary pulmonary nodule † 0.7515 0.8121 0.7152 0.8000 0.7515 0.7697 0.7697 0.8909 0.7455 0.6545 Empyema* 0.8689 0.9344 0.8852 0.8525 0.7869 0.8689 0.8689 0.9672 0.9508 0.9016 Mediastinal widening 0.7495 0.6673 0.8150 0.8224 0.7701 0.7140 0.7682 0.8187 0.7607 0.8579 Chest tube 0.8490 0.8921 0.9461 0.9384 0.9076 0.8983 0.8382 0.9168 0.9461 0.9307 Malignant nodules* 0.8230 0.8230 0.9027 0.8230 0.7965 0.7699 0.7876 0.8230 0.8407 0.8938 Pacemaker 0.9592 0.9796 0.9833 0.9870 0.9852 0.8386 0.7978 0.9870 0.9907 0.9907 Cancer 0.7162 0.7582 0.8331 0.7832 0.7622 0.7595 0.8265 0.8108 0.8384 0.8357 Tracheostomy tube † 0.8688 0.8688 0.9462 0.9634 0.9140 0.8559 0.7892 0.9548 0.9570 0.9720 Atrial dilation* 0.7931 0.6897 0.7241 0.7241 0.6552 0.3793 0.9828 0.7069 0.7759 0.8276 MPA dilation † 0.6535 0.7008 0.8110 0.7874 0.7559 0.7559 0.5748 0.8346 0.7480 0.7795 Emphysema † 0.8962 0.7877 0.8750 0.8750 0.8585 0.8396 0.7123 0.8821 0.8349 0.8962 Interstitial edema 0.8768 0.8594 0.8272 0.8346 0.8456 0.9035 0.7914 0.8851 0.8483 0.8171 Diaphragmatic hernia † 0.7413 0.8297 0.8391 0.8328 0.7603 0.6719 0.8801 0.8486 0.8360 0.8454 Cavitary nodule* 0.7627 0.8305 0.8983 0.8983 0.7119 0.8305 0.6610 0.8136 0.8644 0.9153 Pericardial effusion † 0.6951 0.7195 0.8293 0.7683 0.8293 0.8841 0.7073 0.8598 0.8902 0.8293 Healed rib fracture † 0.6943 0.5886 0.8086 0.8343 0.7314 0.6771 0.7800 0.8171 0.7943 0.8086 Alveolar edema † 0.8354 0.9304 0.9051 0.8481 0.8418 0.8165 0.7785 0.8291 0.8291 0.8987 Aortic valve replacement* 0.9036 0.9398 0.9639 0.9398 0.8916 0.9398 0.8193 0.9036 0.9277 0.9759 Hilar enlargement 0.7042 0.7967 0.7151 0.7949 0.7840 0.6588 0.7659 0.7858 0.7151 0.7532 Hiatal hernia* 0.7273 0.8333 0.8788 0.8788 0.7879 0.8636 0.9697 0.8788 0.8939 0.9242 Bronchovascular crowding* 0.7714 0.8000 0.8381 0.7429 0.8190 0.7714 0.7429 0.8762 0.8095 0.8571 Internal jugular line † 0.8974 0.9475 0.9761 0.9666 0.9236 0.9260 0.7709 0.9594 0.9666 0.9809 Normal 0.8509 0.8737 0.8777 0.8742 0.8667 0.8440 0.8326 0.8600 0.8754 0.8475 Diffused nodules* 0.8362 0.7155 0.9052 0.8707 0.7586 0.7586 0.7672 0.7931 0.7845 0.8534 Calcified nodules* 0.9273 0.6727 0.7273 0.7636 0.5818 0.8000 0.8364 0.6727 0.7273 0.7091 Bibasilar opacities † 0.7832 0.8182 0.7902 0.7552 0.6853 0.7832 0.8322 0.8112 0.8462 0.8601 Clavicle fracture* 0.8115 0.5738 0.6148 0.5902 0.7295 0.8115 0.5492 0.6967 0.7295 0.5574 Nasogastric tube 0.9752 0.9752 0.9797 0.9955 0.9842 0.9744 0.9285 0.9902 0.9910 0.9947 Central venous line † 0.8633 0.9226 0.9294 0.9271 0.8884 0.7745 0.7585 0.9226 0.9362 0.8975 Left subclavian line* 0.9367 0.9114 0.9367 0.9494 0.7975 0.8987 0.7215 0.8734 0.8987 0.9494 Pneumomediastinum* 0.6400 0.6133 0.8933 0.8933 0.6667 0.6267 0.1867 0.7067 0.7333 0.7733 Interstitial thickening † 0.8088 0.7959 0.8140 0.8114 0.7726 0.7416 0.8450 0.8398 0.8243 0.7494 Dobhoff tube † 0.9327 0.9279 0.9279 0.9808 0.9663 0.8990 0.9567 0.9760 0.9567 0.9712 Bronchiectasis* 0.7662 0.8831 0.8312 0.8442 0.8182 0.7792 0.7532 0.7273 0.8312 0.9481 Swan-Ganz catheter † 0.9372 0.9324 0.9565 0.9420 0.9179 0.9275 0.8116 0.9517 0.9710 0.9710 Bullae* 0.5574 0.8361 0.8033 0.8033 0.8361 0.6721 1.0000 0.8525 0.8033 0.8689 Adenocarcinoma* 0.6731 0.8269 0.7885 0.7692 0.8654 0.7308 0.9808 0.8462 0.8654 0.7885 Mediastinal shift † 0.7989 0.8370 0.8587 0.9076 0.8478 0.8207 0.7554 0.9022 0.8859 0.8696 Mediastinal clips* 0.9174 0.9256 0.9917 0.9669 0.9587 0.9339 0.8347 0.9752 0.9504 0.9752 Opacity 0.7890 0.8434 0.7979 0.7673 0.7942 0.7785 0.7808 0.7658 0.7927 0.7450 Atelectasis 0.7351 0.7431 0.8046 0.8012 0.7599 0.7861 0.7644 0.7760 0.7976 0.7926 Hyperinflation 0.8410 0.8285 0.8420 0.8524 0.8337 0.8347 0.7983 0.8857 0.8462 0.8472 Diffuse fibrosis* 0.9508 0.9672 0.9672 0.9672 1.0000 0.9344 0.8197 0.9672 0.9836 0.9672 Surgical clips 0.4989 0.6518 0.7707 0.7219 0.7219 0.6539 0.7749 0.6985 0.7877 0.8323 Low lung volumes 0.7881 0.7602 0.7849 0.7985 0.8175 0.7255 0.8253 0.7838 0.8074 0.8185 Aortic aneurysm* 0.8952 0.5429 0.6381 0.8857 0.7238 0.3714 0.9524 0.7905 0.6000 0.6095 Pneumonia 0.7414 0.7885 0.7646 0.7708 0.7028 0.7680 0.7728 0.8046 0.7602 0.7667 Consolidation 0.7961 0.8431 0.8122 0.8421 0.7850 0.8075 0.7020 0.8116 0.7843 0.8425 Free subdiaphragmatic air* 0.6735 0.8469 0.8367 0.8469 0.7551 0.7653 0.9388 0.7347 0.8163 0.8571 Continued on next page 68 Abnormality EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BioViL-T Med-CLIP Med-AR-2B (ours) Med-AR-8B (ours) Thoracic spine degenerative changes* 0.8785 0.7850 0.7664 0.8318 0.8411 0.8785 0.7570 0.8224 0.8785 0.8879 Port-a-cath † 0.7187 0.8412 0.9638 0.9471 0.9415 0.7382 0.8162 0.9248 0.9777 0.9861 Fibrosis 0.7458 0.8246 0.8473 0.7745 0.7542 0.7733 0.6313 0.8067 0.8365 0.8580 Lung mass † 0.7228 0.7636 0.8804 0.8533 0.7745 0.6902 0.7772 0.8451 0.8560 0.8370 Endotracheal tube 0.9787 0.9813 0.9750 0.9849 0.9802 0.9704 0.9496 0.9782 0.9818 0.9854 Pulmonary arterial hypertension signs † 0.7600 0.7120 0.6560 0.7440 0.8400 0.7920 0.6640 0.6560 0.6920 0.8240 Effusion 0.8429 0.8601 0.8669 0.8494 0.8336 0.8678 0.7951 0.8414 0.8590 0.8744 Calcification 0.7545 0.7243 0.8220 0.8062 0.7624 0.7294 0.7897 0.7732 0.7947 0.8270 Degenerative spine changes † 0.7630 0.8519 0.8593 0.9333 0.8000 0.8593 0.3259 0.9259 0.9185 0.8963 69 S5.4 SpecificityTable S10: Per-label Specificity comparison across baseline, contrastive, and autoregressive models on the MIMIC dataset. Tail abnormalities are marked with an asterisk (*), and medium-prevalence abnormalities are marked with a dagger ( † ). For each row, the best value is shown in bold and the second-best value is underlined. Abnormality EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BioViL-T Med-CLIP Med-AR-2B (ours) Med-AR-8B (ours) Diaphragm elevation 0.7651 0.7826 0.7739 0.8533 0.8421 0.5545 0.2131 0.8440 0.9018 0.8586 Parenchymal opacities † 0.7392 0.7465 0.8114 0.7207 0.7516 0.7588 0.6673 0.7963 0.7685 0.8239 Nodules 0.6128 0.7120 0.8134 0.7440 0.6781 0.6684 0.5918 0.8142 0.7951 0.7553 Pleural effusion 0.8140 0.8133 0.8389 0.8584 0.7923 0.8009 0.7095 0.8325 0.8493 0.8377 Congestive heart failure* 0.7424 0.7545 0.7655 0.6752 0.8219 0.5473 0.4684 0.7504 0.7846 0.7823 Support devices 0.7846 0.8139 0.8325 0.8189 0.8283 0.7751 0.7241 0.8077 0.8251 0.8110 PICC line 0.6081 0.7148 0.9348 0.9337 0.9065 0.5338 0.5221 0.8650 0.9320 0.9350 Pleural thickening 0.7081 0.6473 0.7937 0.8046 0.7616 0.7330 0.3895 0.7374 0.7692 0.7650 Soft tissue swelling † 0.8486 0.9401 0.9206 0.9586 0.8564 0.8323 0.3187 0.9388 0.9007 0.9167 Neoplastic nodules* 0.7633 0.8222 0.9142 0.9039 0.9316 0.7886 0.1045 0.7535 0.8911 0.9106 Hemothorax* 0.7609 0.7517 0.8081 0.8578 0.8450 0.5879 0.4819 0.8518 0.8375 0.8835 CABG † 0.8938 0.8450 0.8118 0.8687 0.8671 0.8816 0.4189 0.8329 0.8214 0.8368 Scarring* 0.5164 0.6971 0.7053 0.7956 0.6177 0.5741 0.2735 0.7297 0.6508 0.7658 Aortic tortuosity 0.7003 0.7497 0.7734 0.7805 0.7468 0.6102 0.6842 0.8065 0.7728 0.7908 Solitary nodules † 0.6878 0.6143 0.7625 0.7465 0.7611 0.5525 0.5337 0.7808 0.7484 0.8186 Tracheal deviation* 0.5903 0.5326 0.5942 0.6777 0.7270 0.3436 0.2655 0.5713 0.6279 0.6431 Scoliosis † 0.7702 0.6604 0.7788 0.9085 0.8497 0.6189 0.6269 0.8794 0.8406 0.8277 Pulmonary edema 0.7530 0.7881 0.7769 0.7738 0.7864 0.7905 0.6752 0.7741 0.7841 0.7686 Rib fractures 0.6340 0.6834 0.8215 0.7343 0.7847 0.5720 0.2848 0.8017 0.8476 0.8738 ARDS* 0.8507 0.8315 0.8911 0.9165 0.8683 0.8764 0.7415 0.8159 0.8627 0.9179 Aspiration † 0.5870 0.5767 0.5968 0.7129 0.6659 0.4923 0.3466 0.6890 0.6814 0.6658 Flattened diaphragm † 0.7950 0.8115 0.8115 0.8676 0.8697 0.7969 0.5344 0.8046 0.8899 0.8572 Reticular opacities † 0.7315 0.7146 0.8336 0.8590 0.8211 0.7922 0.3387 0.7737 0.7364 0.8427 Kyphosis* 0.8499 0.7766 0.7186 0.8877 0.7846 0.7650 0.3191 0.8704 0.8751 0.9035 Vascular redistribution cephalization 0.5603 0.6453 0.6679 0.6445 0.5916 0.6403 0.4654 0.5857 0.6025 0.6325 Degenerative changes † 0.6020 0.6811 0.6579 0.6870 0.6619 0.6363 0.4163 0.6523 0.6477 0.6375 Dialysis catheter* 0.7521 0.8832 0.8959 0.9418 0.9258 0.4998 0.0325 0.9079 0.9250 0.9455 Pigtail catheter* 0.6281 0.6680 0.9515 0.8651 0.8667 0.6300 0.5015 0.8854 0.9533 0.9574 Hilar lymphadenopathy † 0.6161 0.8011 0.7192 0.7857 0.7579 0.5807 0.2793 0.6891 0.7640 0.8355 Chest tube removal* 0.7331 0.7029 0.8521 0.7975 0.8661 0.3142 0.6211 0.7715 0.8362 0.9171 Retrocardiac opacity † 0.5999 0.5495 0.7234 0.7377 0.4655 0.5877 0.4698 0.6818 0.6462 0.7484 Pneumothorax 0.7576 0.7783 0.9288 0.9003 0.8371 0.7297 0.6483 0.8586 0.8880 0.9096 Vascular congestion † 0.6110 0.5635 0.6165 0.6334 0.6305 0.6044 0.6120 0.5939 0.5117 0.6223 Esophageal drainage tube* 0.7690 0.7940 0.8145 0.7987 0.7952 0.7890 0.6465 0.8078 0.8214 0.8128 Perihilar opacities 0.6700 0.7118 0.6964 0.7079 0.6918 0.6942 0.5642 0.7816 0.7722 0.7511 Interstitial lung disease pattern 0.6311 0.6629 0.7429 0.7351 0.8077 0.7182 0.4358 0.7566 0.7196 0.7621 Extubation* 0.4861 0.7078 0.6261 0.7550 0.6998 0.7180 0.5285 0.7209 0.7375 0.8153 Sternotomy wires 0.8919 0.9095 0.9019 0.9058 0.9106 0.8822 0.5645 0.9084 0.9062 0.9062 MPA enlargement † 0.4746 0.6539 0.6304 0.7763 0.7478 0.7267 0.4046 0.7984 0.8406 0.8868 Subcutaneous emphysema † 0.8549 0.9194 0.9603 0.9476 0.9245 0.7863 0.4162 0.9388 0.9582 0.9623 Bronchial wall thickening* 0.6969 0.7611 0.7577 0.7775 0.5335 0.5394 0.5648 0.7448 0.6565 0.6937 Chest wall deformity † 0.8057 0.7055 0.8152 0.7587 0.7536 0.5866 0.4014 0.7182 0.7694 0.7966 Feeding tube 0.7983 0.8040 0.8182 0.8170 0.8162 0.8112 0.7971 0.8146 0.8284 0.8204 Fluid overload* 0.7029 0.5939 0.5637 0.6128 0.6454 0.5800 0.5629 0.5557 0.6929 0.7470 Sternotomy † 0.8791 0.8909 0.8916 0.8937 0.8902 0.8746 0.3595 0.8948 0.8958 0.8958 Continued on next page 70 Abnormality EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BioViL-T Med-CLIP Med-AR-2B (ours) Med-AR-8B (ours) Granuloma † 0.5792 0.6937 0.8867 0.6844 0.6965 0.6996 0.4879 0.7799 0.8049 0.8232 Mediastinal mass* 0.4966 0.6780 0.7992 0.8578 0.7818 0.4914 0.2051 0.8318 0.8530 0.8062 Artifacts 0.5887 0.6316 0.5790 0.6332 0.5994 0.5350 0.5055 0.6048 0.6973 0.6869 Spinal hardware* 0.8491 0.6141 0.9187 0.9226 0.7689 0.8303 0.2548 0.8173 0.9547 0.9641 Cardiomegaly 0.7150 0.6962 0.7595 0.7887 0.7628 0.6762 0.4889 0.7613 0.7380 0.7786 Solitary pulmonary nodule † 0.7156 0.6948 0.7804 0.6774 0.6524 0.6851 0.5237 0.5490 0.7196 0.8267 Empyema* 0.7212 0.7564 0.8239 0.9065 0.8440 0.6167 0.5265 0.8394 0.8472 0.8950 Mediastinal widening 0.7212 0.7840 0.7677 0.7564 0.7342 0.6845 0.5364 0.7666 0.8274 0.7591 Chest tube 0.7712 0.8578 0.9320 0.8917 0.8761 0.6564 0.5401 0.8804 0.9062 0.9399 Malignant nodules* 0.6292 0.7887 0.8920 0.9020 0.8267 0.7172 0.4033 0.8553 0.9029 0.8548 Pacemaker 0.8986 0.9497 0.9437 0.9515 0.9420 0.8386 0.5273 0.9477 0.9492 0.9500 Cancer 0.7705 0.7807 0.7794 0.8346 0.8083 0.6525 0.3522 0.7963 0.7995 0.8088 Tracheostomy tube † 0.7711 0.8982 0.9550 0.9792 0.9406 0.6456 0.6518 0.9582 0.9860 0.9786 Atrial dilation* 0.6177 0.6702 0.7099 0.6905 0.7088 0.7534 0.0224 0.6529 0.6804 0.6208 MPA dilation † 0.7098 0.6975 0.6513 0.7710 0.6930 0.5819 0.5708 0.7576 0.8247 0.8697 Emphysema † 0.7087 0.8441 0.8135 0.8324 0.8102 0.7780 0.6731 0.8232 0.8606 0.8242 Interstitial edema 0.6327 0.6636 0.7452 0.7297 0.6804 0.6191 0.5698 0.6742 0.7107 0.7305 Diaphragmatic hernia † 0.7305 0.6515 0.7829 0.8803 0.7861 0.7100 0.2305 0.7264 0.8294 0.8896 Cavitary nodule* 0.7799 0.8737 0.7724 0.7957 0.8635 0.6949 0.6579 0.8737 0.7841 0.8025 Pericardial effusion † 0.7924 0.7681 0.7873 0.8483 0.7598 0.5552 0.5285 0.7799 0.7351 0.8224 Healed rib fracture † 0.6361 0.7888 0.8664 0.8094 0.6717 0.6569 0.3744 0.8383 0.8065 0.8939 Alveolar edema † 0.7846 0.7183 0.7762 0.8162 0.8158 0.8332 0.7166 0.8584 0.8166 0.7673 Aortic valve replacement* 0.8836 0.8573 0.8674 0.8636 0.8622 0.8689 0.2774 0.9028 0.8747 0.8890 Hilar enlargement 0.6341 0.6285 0.7790 0.7267 0.6832 0.7096 0.3793 0.7372 0.8012 0.7888 Hiatal hernia* 0.7909 0.8099 0.8088 0.9255 0.8486 0.6836 0.3774 0.8017 0.9030 0.9486 Bronchovascular crowding* 0.7926 0.8222 0.8160 0.8719 0.7811 0.7340 0.4688 0.7864 0.8331 0.7784 Internal jugular line † 0.7263 0.7283 0.8309 0.8650 0.9004 0.6754 0.7166 0.8483 0.8825 0.8808 Normal 0.8467 0.8409 0.8563 0.8648 0.8340 0.8547 0.7952 0.8712 0.8599 0.8830 Diffused nodules* 0.6552 0.8729 0.8631 0.8333 0.8756 0.8253 0.4406 0.8906 0.8981 0.8928 Calcified nodules* 0.2467 0.6020 0.8367 0.7368 0.7654 0.3156 0.2849 0.8671 0.8457 0.8924 Bibasilar opacities † 0.6381 0.6300 0.7090 0.7161 0.7438 0.5763 0.4184 0.6904 0.6267 0.6511 Clavicle fracture* 0.4042 0.6512 0.7543 0.8076 0.6478 0.4451 0.5716 0.8108 0.6398 0.8392 Nasogastric tube 0.8339 0.8486 0.8567 0.8491 0.8411 0.8421 0.8293 0.8462 0.8501 0.8490 Central venous line † 0.5697 0.7235 0.7776 0.7972 0.7941 0.5932 0.5618 0.7818 0.8142 0.8476 Left subclavian line* 0.6483 0.7623 0.8198 0.7979 0.9106 0.7293 0.4635 0.8040 0.9151 0.9163 Pneumomediastinum* 0.8616 0.8535 0.7878 0.7801 0.9146 0.7938 0.8492 0.8847 0.8996 0.8779 Interstitial thickening † 0.7352 0.7616 0.8182 0.7997 0.7770 0.7808 0.4283 0.7658 0.7841 0.8489 Dobhoff tube † 0.8220 0.8903 0.9465 0.8771 0.8787 0.8840 0.7629 0.8469 0.9279 0.9456 Bronchiectasis* 0.8791 0.7786 0.8703 0.8418 0.7981 0.8311 0.5779 0.9152 0.8653 0.7630 Swan-Ganz catheter † 0.8044 0.8897 0.9506 0.9309 0.9621 0.7473 0.4937 0.9063 0.9580 0.9901 Bullae* 0.8401 0.7448 0.7778 0.8233 0.7150 0.7628 0.0094 0.7720 0.7893 0.8543 Adenocarcinoma* 0.6782 0.6383 0.7936 0.7488 0.7157 0.5888 0.0527 0.7796 0.7262 0.8163 Mediastinal shift † 0.8228 0.8037 0.8378 0.8120 0.8335 0.6810 0.4550 0.8192 0.8465 0.9025 Mediastinal clips* 0.9291 0.8959 0.8797 0.9015 0.9070 0.8848 0.2924 0.8936 0.9297 0.8912 Opacity 0.5553 0.5286 0.6269 0.6359 0.5824 0.5851 0.4921 0.6212 0.5935 0.6649 Atelectasis 0.7054 0.7212 0.7100 0.7011 0.7065 0.6508 0.5465 0.7160 0.7056 0.7174 Hyperinflation 0.8058 0.8290 0.8232 0.8300 0.8345 0.8050 0.6565 0.7952 0.8356 0.8423 Diffuse fibrosis* 0.8818 0.9116 0.9415 0.9332 0.9180 0.8988 0.5378 0.9718 0.9443 0.9686 Surgical clips 0.7848 0.6267 0.7728 0.7229 0.7721 0.6119 0.4224 0.7331 0.8260 0.8720 Low lung volumes 0.7491 0.7801 0.7707 0.7693 0.7433 0.7613 0.4979 0.7825 0.7612 0.7620 Aortic aneurysm* 0.2585 0.8244 0.8438 0.6599 0.6752 0.7435 0.1242 0.6889 0.8942 0.9367 Pneumonia 0.6317 0.6396 0.7399 0.7157 0.7392 0.6325 0.4575 0.6633 0.7198 0.7282 Consolidation 0.6345 0.6394 0.7232 0.6761 0.6886 0.6401 0.6120 0.6955 0.7255 0.6783 Free subdiaphragmatic air* 0.7240 0.6598 0.7074 0.7410 0.8481 0.6010 0.1877 0.8667 0.8422 0.9123 Continued on next page 71 Abnormality EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BioViL-T Med-CLIP Med-AR-2B (ours) Med-AR-8B (ours) Thoracic spine degenerative changes* 0.6473 0.6871 0.7707 0.6768 0.6098 0.5664 0.4928 0.6985 0.6467 0.6778 Port-a-cath † 0.7700 0.8969 0.9303 0.9282 0.9349 0.5169 0.2368 0.9518 0.9560 0.9664 Fibrosis 0.7149 0.6950 0.7744 0.8397 0.7682 0.6766 0.6705 0.7982 0.7778 0.7715 Lung mass † 0.7304 0.8325 0.7803 0.8324 0.8314 0.7063 0.3664 0.8075 0.8062 0.8480 Endotracheal tube 0.8955 0.9068 0.9231 0.9262 0.9170 0.8978 0.8704 0.9226 0.9375 0.9387 Pulmonary arterial hypertension signs † 0.5727 0.7024 0.7740 0.7729 0.5885 0.5924 0.5498 0.8519 0.8143 0.7156 Effusion 0.8233 0.8244 0.8444 0.8618 0.8333 0.8035 0.6984 0.8560 0.8423 0.8305 Calcification 0.6772 0.7240 0.7043 0.7675 0.7165 0.6506 0.4755 0.8084 0.7945 0.7883 Degenerative spine changes † 0.6893 0.6877 0.7054 0.6631 0.7214 0.5777 0.7146 0.6375 0.6735 0.6978 72 S5.5 Youden’s J Table S11: Per-label Youden’s J Score comparison across baseline, contrastive, and autoregressive models on the MIMIC dataset. Tail abnormalities are marked with an asterisk (*), and medium-prevalence abnormalities are marked with a dagger ( † ). For each row, the best value is shown in bold and the second-best value is underlined. Abnormality EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BioViL-T Med-CLIP Med-AR-2B (ours) Med-AR-8B (ours) Diaphragm elevation 0.4764 0.5133 0.5681 0.6641 0.5976 0.2865 0.1302 0.6617 0.6905 0.7108 Parenchymal opacities † 0.6015 0.6387 0.6916 0.6429 0.6438 0.5732 0.4278 0.6406 0.6487 0.6562 Nodules 0.3323 0.3817 0.5002 0.4714 0.4016 0.3460 0.1449 0.4381 0.4596 0.5024 Pleural effusion 0.6656 0.6842 0.7136 0.7122 0.6658 0.6715 0.4940 0.6996 0.7028 0.7042 Congestive heart failure* 0.5519 0.5799 0.6226 0.5482 0.5838 0.4679 0.3890 0.6393 0.6259 0.6394 Support devices 0.7273 0.7440 0.7841 0.7687 0.7584 0.7160 0.6721 0.7647 0.7893 0.7859 PICC line 0.3894 0.6108 0.8943 0.8875 0.8208 0.3199 0.3034 0.7446 0.8886 0.8974 Pleural thickening 0.3805 0.4439 0.5420 0.5460 0.4392 0.3468 0.1740 0.5288 0.5571 0.5650 Soft tissue swelling † 0.5441 0.6326 0.7116 0.7108 0.6564 0.4860 0.1635 0.7000 0.7216 0.7316 Neoplastic nodules* 0.5237 0.6139 0.8413 0.7685 0.6608 0.5074 0.0837 0.6910 0.7661 0.7960 Hemothorax* 0.5749 0.6354 0.7383 0.6718 0.7171 0.4251 0.3540 0.7123 0.7794 0.7440 CABG † 0.6382 0.7450 0.7192 0.7576 0.7190 0.5964 0.1559 0.7440 0.7547 0.7738 Scarring* 0.4005 0.4217 0.5314 0.5492 0.4003 0.4147 0.0561 0.5268 0.5638 0.5339 Aortic tortuosity 0.4880 0.5274 0.5722 0.6095 0.5466 0.4492 0.2335 0.6053 0.6018 0.6298 Solitary nodules † 0.4396 0.4683 0.5216 0.4837 0.4399 0.4722 0.3877 0.4961 0.5148 0.4755 Tracheal deviation* 0.3830 0.3497 0.3991 0.3240 0.2880 0.2948 0.1801 0.4006 0.3840 0.4236 Scoliosis † 0.4906 0.4663 0.5091 0.6815 0.5504 0.3722 0.2124 0.6261 0.6038 0.6994 Pulmonary edema 0.6175 0.6233 0.6591 0.6516 0.6205 0.6195 0.4909 0.6465 0.6505 0.6540 Rib fractures 0.2864 0.3581 0.6160 0.5802 0.3755 0.2535 0.1187 0.5859 0.5856 0.6306 ARDS* 0.7683 0.7491 0.8087 0.7518 0.7507 0.8058 0.5886 0.7688 0.7686 0.7767 Aspiration † 0.3278 0.3591 0.4232 0.4488 0.3749 0.3089 0.2121 0.4298 0.4296 0.4458 Flattened diaphragm † 0.6867 0.7160 0.7223 0.7275 0.7168 0.6695 0.3815 0.7345 0.7116 0.7489 Reticular opacities † 0.5254 0.5848 0.6733 0.6376 0.5845 0.5327 0.1479 0.6592 0.6601 0.6748 Kyphosis* 0.6089 0.6682 0.6704 0.6467 0.6521 0.5843 0.0902 0.7017 0.7185 0.6746 Vascular redistribution cephalization 0.4254 0.4386 0.4762 0.4670 0.4470 0.4344 0.3048 0.4562 0.4596 0.4728 Degenerative changes † 0.4546 0.4460 0.4719 0.4975 0.4935 0.3907 0.1181 0.4734 0.5003 0.5042 Dialysis catheter* 0.4558 0.7351 0.8218 0.8183 0.7653 0.2652 0.0202 0.8215 0.8509 0.9085 Pigtail catheter* 0.4923 0.5445 0.8713 0.7910 0.7309 0.4386 0.2484 0.7743 0.8484 0.8895 Hilar lymphadenopathy † 0.3402 0.4218 0.5468 0.5295 0.4919 0.3245 0.0724 0.5019 0.5177 0.6040 Chest tube removal* 0.5002 0.6070 0.6740 0.7016 0.6058 0.2731 0.1416 0.6619 0.7266 0.6979 Retrocardiac opacity † 0.3373 0.3260 0.4273 0.4249 0.3705 0.2749 0.2631 0.4025 0.3725 0.4579 Pneumothorax 0.5679 0.6453 0.8164 0.7642 0.6928 0.5503 0.3132 0.7370 0.7633 0.7941 Vascular congestion † 0.3792 0.4180 0.4801 0.4334 0.3987 0.4180 0.3438 0.4348 0.4390 0.4678 Esophageal drainage tube* 0.7310 0.7940 0.7892 0.7860 0.7319 0.7637 0.3554 0.7698 0.7961 0.8128 Perihilar opacities 0.4381 0.4622 0.5283 0.5132 0.4883 0.4446 0.2969 0.4949 0.5138 0.5405 Interstitial lung disease pattern 0.3942 0.4278 0.5021 0.4935 0.4305 0.4036 0.1820 0.4902 0.4914 0.5096 Extubation* 0.4724 0.5434 0.5165 0.4673 0.5080 0.5262 0.3093 0.5839 0.6827 0.6509 Sternotomy wires 0.8566 0.8870 0.8971 0.9010 0.8865 0.8421 0.2740 0.8972 0.8998 0.9030 MPA enlargement † 0.3069 0.3620 0.4130 0.5465 0.4186 0.3168 0.1996 0.5065 0.4990 0.5949 Subcutaneous emphysema † 0.6375 0.7703 0.8982 0.8855 0.7941 0.5192 0.2485 0.8518 0.8588 0.9002 Bronchial wall thickening* 0.3826 0.3801 0.4815 0.4537 0.3049 0.2727 0.1934 0.4496 0.4470 0.4366 Chest wall deformity † 0.3996 0.4314 0.5512 0.5658 0.4135 0.3328 0.1933 0.5811 0.5968 0.6342 Feeding tube 0.7682 0.7866 0.8055 0.8123 0.8020 0.7843 0.7148 0.8051 0.8173 0.8141 Fluid overload* 0.4551 0.4523 0.5018 0.4358 0.4684 0.4650 0.3771 0.4938 0.4982 0.4904 Sternotomy † 0.8292 0.8857 0.8811 0.8885 0.8771 0.8169 0.2046 0.8817 0.8827 0.8853 Granuloma † 0.3811 0.3352 0.5376 0.4627 0.3144 0.2845 0.2568 0.5063 0.5643 0.6062 Continued on next page 73 Abnormality EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BioViL-T Med-CLIP Med-AR-2B (ours) Med-AR-8B (ours) Mediastinal mass* 0.3416 0.3912 0.5511 0.5090 0.4407 0.2433 0.0656 0.5217 0.5507 0.5891 Artifacts 0.2976 0.2825 0.3430 0.3471 0.3398 0.2292 0.1840 0.3373 0.3669 0.3644 Spinal hardware* 0.2589 0.2043 0.7712 0.5947 0.4246 0.1582 0.1728 0.5058 0.6924 0.7510 Cardiomegaly 0.5258 0.5073 0.5719 0.5683 0.5415 0.4810 0.3195 0.5611 0.5718 0.5757 Solitary pulmonary nodule † 0.4671 0.5069 0.4956 0.4774 0.4039 0.4548 0.2934 0.4399 0.4651 0.4812 Empyema* 0.5901 0.6908 0.7091 0.7590 0.6309 0.4856 0.3954 0.8066 0.7980 0.7966 Mediastinal widening 0.4707 0.4513 0.5827 0.5788 0.5043 0.3985 0.3046 0.5853 0.5881 0.6170 Chest tube 0.6202 0.7499 0.8781 0.8301 0.7837 0.5547 0.3783 0.7972 0.8523 0.8706 Malignant nodules* 0.4522 0.6117 0.7947 0.7250 0.6232 0.4871 0.1909 0.6783 0.7436 0.7486 Pacemaker 0.8578 0.9293 0.9270 0.9385 0.9272 0.6772 0.3251 0.9347 0.9399 0.9407 Cancer 0.4867 0.5389 0.6125 0.6178 0.5705 0.4120 0.1787 0.6071 0.6379 0.6445 Tracheostomy tube † 0.6399 0.7670 0.9012 0.9426 0.8546 0.5015 0.4410 0.9130 0.9430 0.9506 Atrial dilation* 0.4108 0.3599 0.4340 0.4146 0.3640 0.1327 0.0052 0.3598 0.4563 0.4484 MPA dilation † 0.3633 0.3983 0.4623 0.5584 0.4489 0.3378 0.1456 0.5922 0.5727 0.6492 Emphysema † 0.6049 0.6318 0.6885 0.7074 0.6687 0.6176 0.3854 0.7053 0.6955 0.7204 Interstitial edema 0.5095 0.5230 0.5724 0.5643 0.5260 0.5226 0.3612 0.5593 0.5590 0.5476 Diaphragmatic hernia † 0.4718 0.4812 0.6220 0.7131 0.5464 0.3819 0.1106 0.5750 0.6654 0.7350 Cavitary nodule* 0.5426 0.7042 0.6707 0.6940 0.5754 0.5254 0.3189 0.7015 0.6485 0.7178 Pericardial effusion † 0.4875 0.4876 0.6166 0.6166 0.5891 0.4393 0.2358 0.6397 0.6253 0.6517 Healed rib fracture † 0.3304 0.3774 0.6750 0.6437 0.4031 0.3340 0.1544 0.6554 0.6008 0.7025 Alveolar edema † 0.6200 0.6487 0.6813 0.6643 0.6576 0.6497 0.4951 0.6875 0.6457 0.6660 Aortic valve replacement* 0.7872 0.7971 0.8313 0.8034 0.7538 0.8087 0.0967 0.8064 0.8024 0.8649 Hilar enlargement 0.3383 0.4252 0.4941 0.5216 0.4672 0.3684 0.1452 0.5230 0.5163 0.5420 Hiatal hernia* 0.5182 0.6432 0.6876 0.8043 0.6365 0.5472 0.3471 0.6805 0.7969 0.8728 Bronchovascular crowding* 0.5640 0.6222 0.6541 0.6148 0.6001 0.5054 0.2117 0.6626 0.6426 0.6355 Internal jugular line † 0.6237 0.6758 0.8070 0.8316 0.8240 0.6014 0.4875 0.8077 0.8491 0.8617 Normal 0.6976 0.7146 0.7340 0.7390 0.7007 0.6987 0.6278 0.7312 0.7353 0.7305 Diffused nodules* 0.4914 0.5884 0.7683 0.7040 0.6342 0.5839 0.2078 0.6837 0.6826 0.7462 Calcified nodules* 0.1740 0.2747 0.5640 0.5004 0.3472 0.1156 0.1213 0.5398 0.5730 0.6015 Bibasilar opacities † 0.4213 0.4482 0.4992 0.4713 0.4291 0.3595 0.2506 0.5016 0.4729 0.5112 Clavicle fracture* 0.2157 0.2250 0.3691 0.3978 0.3773 0.2566 0.1208 0.5075 0.3693 0.3966 Nasogastric tube 0.8091 0.8238 0.8364 0.8446 0.8253 0.8165 0.7578 0.8364 0.8411 0.8437 Central venous line † 0.4330 0.6461 0.7070 0.7243 0.6825 0.3677 0.3203 0.7044 0.7504 0.7451 Left subclavian line* 0.5850 0.6737 0.7565 0.7473 0.7081 0.6280 0.1850 0.6774 0.8138 0.8657 Pneumomediastinum* 0.5016 0.4668 0.6811 0.6734 0.5813 0.4205 0.0359 0.5914 0.6329 0.6512 Interstitial thickening † 0.5440 0.5575 0.6322 0.6111 0.5496 0.5224 0.2733 0.6056 0.6084 0.5983 Dobhoff tube † 0.7547 0.8182 0.8744 0.8579 0.8450 0.7830 0.7196 0.8229 0.8846 0.9168 Bronchiectasis* 0.6453 0.6617 0.7015 0.6860 0.6163 0.6103 0.3311 0.6425 0.6965 0.7111 Swan-Ganz catheter † 0.7416 0.8221 0.9071 0.8729 0.8800 0.6748 0.3053 0.8580 0.9290 0.9611 Bullae* 0.3975 0.5809 0.5811 0.6266 0.5511 0.4349 0.0094 0.6245 0.5926 0.7232 Adenocarcinoma* 0.3513 0.4652 0.5821 0.5180 0.5811 0.3196 0.0335 0.6258 0.5916 0.6048 Mediastinal shift † 0.6217 0.6407 0.6965 0.7196 0.6813 0.5017 0.2104 0.7214 0.7324 0.7721 Mediastinal clips* 0.8465 0.8215 0.8714 0.8684 0.8657 0.8187 0.1271 0.8688 0.8801 0.8664 Opacity 0.3443 0.3720 0.4248 0.4032 0.3766 0.3636 0.2729 0.3870 0.3862 0.4099 Atelectasis 0.4405 0.4643 0.5146 0.5023 0.4664 0.4369 0.3109 0.4920 0.5032 0.5100 Hyperinflation 0.6468 0.6575 0.6652 0.6824 0.6682 0.6397 0.4548 0.6809 0.6818 0.6895 Diffuse fibrosis* 0.8326 0.8788 0.9087 0.9004 0.9180 0.8332 0.3575 0.9390 0.9279 0.9358 Surgical clips 0.2837 0.2785 0.5435 0.4448 0.4940 0.2658 0.1973 0.4316 0.6137 0.7043 Low lung volumes 0.5372 0.5403 0.5556 0.5678 0.5608 0.4868 0.3232 0.5663 0.5686 0.5805 Aortic aneurysm* 0.1537 0.3673 0.4819 0.5456 0.3990 0.1149 0.0766 0.4794 0.4942 0.5462 Pneumonia 0.3731 0.4281 0.5045 0.4865 0.4420 0.4005 0.2303 0.4679 0.4800 0.4949 Consolidation 0.4306 0.4825 0.5354 0.5182 0.4736 0.4476 0.3140 0.5071 0.5098 0.5208 Free subdiaphragmatic air* 0.3975 0.5067 0.5441 0.5879 0.6032 0.3663 0.1265 0.6014 0.6585 0.7694 Thoracic spine degenerative changes* 0.5258 0.4721 0.5371 0.5086 0.4509 0.4449 0.2498 0.5209 0.5252 0.5657 Continued on next page 74 Abnormality EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BioViL-T Med-CLIP Med-AR-2B (ours) Med-AR-8B (ours) Port-a-cath † 0.4887 0.7381 0.8941 0.8753 0.8764 0.2551 0.0530 0.8766 0.9337 0.9525 Fibrosis 0.4607 0.5196 0.6217 0.6142 0.5224 0.4499 0.3018 0.6049 0.6143 0.6295 Lung mass † 0.4532 0.5961 0.6607 0.6857 0.6059 0.3965 0.1436 0.6526 0.6622 0.6850 Endotracheal tube 0.8742 0.8881 0.8981 0.9111 0.8972 0.8682 0.8200 0.9008 0.9193 0.9241 Pulmonary arterial hypertension signs † 0.3327 0.4144 0.4300 0.5169 0.4285 0.3844 0.2138 0.5079 0.5063 0.5396 Effusion 0.6662 0.6845 0.7113 0.7112 0.6669 0.6713 0.4935 0.6974 0.7013 0.7049 Calcification 0.4317 0.4483 0.5263 0.5737 0.4789 0.3800 0.2652 0.5816 0.5892 0.6153 Degenerative spine changes † 0.4523 0.5396 0.5647 0.5964 0.5214 0.4370 0.0405 0.5634 0.5920 0.5941 75 S6 PadChest: Per-Label ResultsThe following tables report the existing results for 29 labels, in the order AUROC, AUPRC, sensitivity, specificity, and Youden’s J . S6.1 AUROCTable S12: Per-label AUROC comparison across baseline, contrastive, and autoregressive models on the PadChest dataset. Tail abnormalities are markedwith an asterisk (*), and medium-prevalence abnormalities are marked with a dagger ( † ). For each row, the best value is shown in bold and the second-best value is underlined. Abnormality EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BioViL-T Med-CLIP Med-AR (ours) Med-AR-8B (ours) NSG Tube † 0.9695 0.9685 0.9925 0.9928 0.9291 0.9888 0.9653 0.9899 0.9934 0.9919 Volume Loss* 0.7783 0.7486 0.9117 0.9137 0.6554 0.8210 0.5479 0.9087 0.9226 0.9163 Kyphosis † 0.8535 0.8203 0.8683 0.9009 0.6513 0.8491 0.5987 0.8914 0.8988 0.9029 Alveolar Pattern † 0.8863 0.8728 0.9278 0.9270 0.7708 0.9055 0.8291 0.9219 0.9277 0.9204 Pacemaker* 0.9851 0.9819 0.9967 0.9954 0.9244 0.9944 0.9641 0.9954 0.9969 0.9955 COPD Signs 0.8174 0.8084 0.8324 0.8386 0.6645 0.8168 0.6061 0.8318 0.8436 0.8407 Infiltrates 0.7314 0.7132 0.8069 0.8068 0.6065 0.7658 0.6513 0.7880 0.8125 0.8001 Vertebral Degenerative Changes † 0.7564 0.7351 0.7897 0.8368 0.6221 0.7892 0.5485 0.8247 0.8484 0.8366 Apical Pleural Thickening* 0.7702 0.7448 0.8659 0.8721 0.6104 0.7656 0.5514 0.8787 0.8966 0.8796 Pleural Effusion 0.9254 0.8994 0.9578 0.9572 0.8145 0.9454 0.8444 0.9539 0.9578 0.9535 Atelectasis* 0.8360 0.8197 0.8725 0.8763 0.7359 0.8547 0.7866 0.8705 0.8814 0.8704 Central Venous Catheter Via Jugular Vein* 0.9646 0.9602 0.9945 0.9934 0.9399 0.9912 0.9806 0.9949 0.9954 0.9926 Air Trapping † 0.8014 0.7936 0.8043 0.8191 0.6285 0.8042 0.6393 0.8151 0.8241 0.8132 Calcified Granuloma* 0.7086 0.7017 0.8730 0.8381 0.6146 0.7148 0.5435 0.8600 0.8880 0.8769 Interstitial Pattern 0.8179 0.7892 0.8712 0.8688 0.7008 0.8442 0.7646 0.8634 0.8747 0.8645 Aortic Elongation 0.8321 0.8050 0.8847 0.8922 0.6674 0.8630 0.6239 0.8874 0.8989 0.8931 Laminar Atelectasis † 0.7137 0.6822 0.8700 0.8526 0.5733 0.7761 0.5932 0.8413 0.8701 0.8572 Costophrenic Angle Blunting † 0.7862 0.7547 0.8714 0.8741 0.6535 0.8522 0.6852 0.8679 0.8781 0.8708 Scoliosis 0.7402 0.7174 0.7963 0.8561 0.5989 0.7522 0.5378 0.8378 0.8699 0.8685 Chronic Changes 0.7846 0.7731 0.8035 0.8060 0.6826 0.7803 0.6157 0.8142 0.8252 0.8192 Pneumonia 0.7651 0.7490 0.8577 0.8516 0.6109 0.8106 0.6395 0.8312 0.8548 0.8415 Callus Rib Fracture* 0.6777 0.6750 0.8425 0.7856 0.6002 0.6830 0.5561 0.8313 0.8575 0.8398 Sternotomy* 0.9769 0.9805 0.9964 0.9950 0.8824 0.9947 0.9747 0.9963 0.9963 0.9964 Unchanged 0.6619 0.6527 0.6867 0.6859 0.5843 0.6618 0.6262 0.6977 0.7075 0.6972 Vascular Hilar Enlargement † 0.7038 0.6887 0.7467 0.7538 0.6272 0.7005 0.6111 0.7598 0.7776 0.7731 Bronchiectasis* 0.7490 0.7332 0.8380 0.8268 0.6346 0.7812 0.6471 0.8230 0.8430 0.8330 Fibrotic Band* 0.7388 0.7107 0.8562 0.8536 0.6428 0.7436 0.5516 0.8427 0.8659 0.8597 Cardiomegaly 0.8851 0.8696 0.9214 0.9237 0.6998 0.8918 0.7164 0.9193 0.9265 0.9191 Nodule † 0.6698 0.6720 0.7887 0.7684 0.5785 0.6894 0.4982 0.7661 0.7885 0.7696 76 S6.2 AUPRCTable S13: Per-label AUPRC comparison across baseline, contrastive, and autoregressive models on the PadChest dataset. Tail abnormalities are markedwith an asterisk (*), and medium-prevalence abnormalities are marked with a dagger ( † ). For each row, the best value is shown in bold and the second-best value is underlined. Abnormality EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BioViL-T Med-CLIP Med-AR-2B (ours) Med-AR-8B (ours) NSG tube † 0.5976 0.5641 0.8669 0.8691 0.4898 0.8268 0.5539 0.8283 0.8728 0.8630 Volume loss* 0.1298 0.1106 0.3903 0.3851 0.0720 0.1751 0.0347 0.3933 0.4516 0.4227 Kyphosis † 0.3222 0.2504 0.3504 0.4490 0.1112 0.2906 0.0687 0.4186 0.4494 0.4390 Alveolar pattern † 0.4612 0.4173 0.5698 0.5612 0.2338 0.5154 0.3237 0.5623 0.5651 0.5504 Pacemaker* 0.7318 0.7209 0.9142 0.9008 0.6551 0.8997 0.5389 0.9164 0.9388 0.9374 COPD signs 0.5282 0.5107 0.5612 0.5695 0.3509 0.5288 0.2584 0.5638 0.5862 0.5796 Infiltrates 0.1607 0.1481 0.2397 0.2463 0.1010 0.1966 0.1115 0.2163 0.2518 0.2418 Vertebral degenerative changes † 0.1326 0.1187 0.1683 0.2387 0.0766 0.1580 0.0542 0.2163 0.2602 0.2461 Apical pleural thickening* 0.1286 0.1088 0.2909 0.2992 0.0707 0.1220 0.0433 0.3374 0.3643 0.3460 Pleural effusion 0.6255 0.5219 0.7803 0.7796 0.3477 0.7169 0.3728 0.7684 0.7824 0.7681 Atelectasis* 0.1504 0.1126 0.2396 0.2413 0.0818 0.1823 0.0992 0.2231 0.2452 0.2325 Central venous catheter via jugular vein* 0.4037 0.3664 0.8416 0.8446 0.3505 0.8060 0.6551 0.8451 0.8682 0.8642 Air trapping † 0.2203 0.2123 0.2206 0.2319 0.0959 0.2158 0.0815 0.2390 0.2401 0.2277 Calcified granuloma* 0.0720 0.0686 0.3702 0.2902 0.0526 0.0702 0.0344 0.3677 0.4253 0.4036 Interstitial pattern 0.3124 0.2650 0.4232 0.4175 0.1858 0.3593 0.2207 0.4112 0.4253 0.4084 Aortic elongation 0.3602 0.3171 0.4735 0.5059 0.1869 0.4347 0.1389 0.4845 0.5202 0.5030 Laminar atelectasis † 0.1174 0.0934 0.4013 0.3733 0.0620 0.2088 0.0594 0.3539 0.4053 0.3913 Costophrenic angle blunting † 0.2182 0.1906 0.4023 0.4045 0.1314 0.3542 0.1214 0.4025 0.4218 0.4071 Scoliosis 0.1914 0.1621 0.3033 0.4324 0.1031 0.2028 0.0768 0.3974 0.4649 0.4664 Chronic changes 0.2159 0.2001 0.2318 0.2422 0.1521 0.2172 0.0961 0.2660 0.2728 0.2617 Pneumonia 0.2601 0.2248 0.4333 0.4231 0.1193 0.3302 0.1259 0.3825 0.4401 0.4118 Callus rib fracture* 0.0517 0.0532 0.3217 0.1570 0.0418 0.0553 0.0317 0.3159 0.3513 0.3249 Sternotomy* 0.6916 0.6944 0.8915 0.8881 0.5077 0.8628 0.5397 0.8883 0.8814 0.8813 Unchanged 0.2034 0.1964 0.2261 0.2258 0.1666 0.2018 0.1876 0.2339 0.2381 0.2370 Vascular hilar enlargement † 0.1017 0.0934 0.1321 0.1344 0.0772 0.0997 0.0637 0.1457 0.1584 0.1570 Bronchiectasis* 0.0864 0.0700 0.1683 0.1621 0.0551 0.1064 0.0561 0.1581 0.1933 0.1861 Fibrotic band* 0.1344 0.1008 0.3190 0.3071 0.0825 0.1262 0.0542 0.3162 0.3526 0.3417 Cardiomegaly 0.6051 0.5630 0.7120 0.7159 0.3205 0.6235 0.2696 0.7041 0.7238 0.7051 Nodule † 0.0608 0.0615 0.1583 0.1305 0.0457 0.0708 0.0322 0.1435 0.1740 0.1548 77 S6.3 SensitivityTable S14: Per-label Sensitivity comparison across baseline, contrastive, and autoregressive models on the PadChest dataset. Tail abnormalities are marked with an asterisk (*), and medium-prevalence abnormalities are marked with a dagger ( † ). For each row, the best value is shown in bold and the second-best value is underlined. Abnormality EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BioViL-T Med-CLIP Med-AR-2B (ours) Med-AR-8B (ours) NSG tube † 0.9554 0.9705 0.9857 0.9777 0.8375 0.9737 0.9450 0.9721 0.9809 0.9785 Volume loss* 0.6830 0.7692 0.8276 0.8382 0.5040 0.7944 0.6631 0.8130 0.8236 0.8554 Kyphosis † 0.8060 0.7610 0.8419 0.8659 0.6247 0.8052 0.8352 0.8479 0.8142 0.8831 Alveolar pattern † 0.8368 0.8318 0.8677 0.8878 0.7549 0.8066 0.7026 0.8601 0.8570 0.8715 Pacemaker* 0.9643 0.9475 0.9855 0.9788 0.7969 0.9777 0.9375 0.9810 0.9900 0.9922 COPD signs 0.7921 0.7526 0.7898 0.7789 0.5750 0.7676 0.7691 0.7855 0.8083 0.7829 Infiltrates 0.7383 0.7278 0.7925 0.7260 0.5320 0.7137 0.5665 0.7525 0.7586 0.7629 Vertebral degenerative changes † 0.8148 0.8131 0.8064 0.7896 0.6563 0.7519 0.9279 0.7611 0.8399 0.8106 Apical pleural thickening* 0.7497 0.7920 0.7584 0.8180 0.5287 0.7443 0.8212 0.7941 0.8256 0.8061 Pleural effusion 0.8464 0.8619 0.9031 0.9157 0.7558 0.8963 0.8007 0.9060 0.9043 0.9022 Atelectasis* 0.7818 0.7629 0.8171 0.8388 0.7290 0.7696 0.7141 0.8279 0.8428 0.8225 Central venous catheter via jugular vein* 0.9697 0.9818 0.9818 0.9806 0.9126 0.9757 0.9575 0.9745 0.9769 0.9636 Air trapping † 0.7697 0.7853 0.8284 0.8187 0.5862 0.8187 0.8692 0.7623 0.8291 0.7801 Calcified granuloma* 0.6493 0.7537 0.7483 0.7309 0.6292 0.7323 0.9277 0.8086 0.8220 0.7403 Interstitial pattern 0.7568 0.7472 0.8347 0.8126 0.6779 0.7784 0.7085 0.8080 0.8246 0.8226 Aortic elongation 0.8629 0.7800 0.8637 0.8523 0.6888 0.8322 0.8731 0.8570 0.8853 0.8695 Laminar atelectasis † 0.7576 0.7732 0.7837 0.7350 0.7915 0.6855 0.8297 0.7515 0.7454 0.7202 Costophrenic angle blunting † 0.7437 0.6718 0.8138 0.7977 0.6391 0.7793 0.6994 0.8144 0.8195 0.8011 Scoliosis 0.7196 0.6700 0.7202 0.7852 0.6468 0.7422 0.9302 0.7763 0.7751 0.8013 Chronic changes 0.8368 0.8049 0.8203 0.8126 0.6698 0.7753 0.8736 0.7907 0.8110 0.8148 Pneumonia 0.6411 0.7501 0.7679 0.7593 0.5947 0.7097 0.6433 0.7453 0.7949 0.7620 Callus rib fracture* 0.7937 0.7831 0.6973 0.7169 0.7154 0.6657 0.6913 0.7229 0.7410 0.7741 Sternotomy* 0.9297 0.9568 0.9932 0.9851 0.7392 0.9770 0.9284 0.9905 0.9959 0.9919 Unchanged 0.7794 0.7473 0.7717 0.6759 0.5058 0.7436 0.6847 0.7324 0.8160 0.6888 Vascular hilar enlargement † 0.7587 0.7708 0.8339 0.7872 0.6298 0.7189 0.7664 0.7517 0.7664 0.7085 Bronchiectasis* 0.8125 0.7217 0.7932 0.7887 0.5417 0.7887 0.5045 0.7812 0.8021 0.7664 Fibrotic band* 0.6467 0.7870 0.7527 0.6899 0.5388 0.7910 0.3052 0.7527 0.7242 0.7939 Cardiomegaly 0.8370 0.8370 0.8749 0.8693 0.6636 0.8813 0.8462 0.8765 0.8886 0.8698 Nodule † 0.6344 0.6370 0.7506 0.6408 0.5659 0.7080 0.9574 0.7080 0.6977 0.7287 78 S6.4 SpecificityTable S15: Per-label Specificity comparison across baseline, contrastive, and autoregressive models on the PadChest dataset. Tail abnormalities are marked with an asterisk (*), and medium-prevalence abnormalities are marked with a dagger ( † ). For each row, the best value is shown in bold and the second-best value is underlined. Abnormality EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BioViL-T Med-CLIP Med-AR-2B (ours) Med-AR-8B (ours) NSG tube † 0.9266 0.9147 0.9654 0.9715 0.8855 0.9590 0.9095 0.9643 0.9723 0.9712 Volume loss* 0.7478 0.5952 0.8366 0.8373 0.7262 0.7140 0.4524 0.8477 0.8657 0.8266 Kyphosis † 0.7370 0.7264 0.7439 0.7791 0.6014 0.7316 0.3773 0.7823 0.8284 0.7720 Alveolar pattern † 0.7680 0.7514 0.8327 0.8112 0.6522 0.8336 0.8243 0.8334 0.8459 0.8142 Pacemaker* 0.9373 0.9372 0.9900 0.9850 0.9082 0.9771 0.8835 0.9887 0.9938 0.9878 COPD signs 0.6898 0.7143 0.7166 0.7400 0.6602 0.7105 0.3865 0.7168 0.7214 0.7369 Infiltrates 0.6214 0.6082 0.6908 0.7482 0.6384 0.6853 0.6577 0.6896 0.7314 0.6992 Vertebral degenerative changes † 0.5695 0.5372 0.6248 0.7164 0.5338 0.6851 0.2207 0.7396 0.6981 0.7126 Apical pleural thickening* 0.6601 0.5633 0.8104 0.7763 0.6465 0.6454 0.2872 0.8035 0.8113 0.7992 Pleural effusion 0.8692 0.7963 0.8928 0.8810 0.7498 0.8692 0.7499 0.8804 0.8906 0.8843 Atelectasis* 0.7844 0.7646 0.7928 0.7745 0.6293 0.8116 0.7398 0.7806 0.7843 0.7801 Central venous catheter via jugular vein* 0.9097 0.9005 0.9685 0.9605 0.8592 0.9484 0.9265 0.9757 0.9853 0.9859 Air trapping † 0.6949 0.6716 0.6300 0.6659 0.6041 0.6402 0.3453 0.7156 0.6691 0.6949 Calcified granuloma* 0.6689 0.5392 0.8242 0.7911 0.5450 0.5867 0.1760 0.7435 0.7888 0.8485 Interstitial pattern 0.7328 0.7040 0.7536 0.7728 0.6211 0.7610 0.6995 0.7622 0.7671 0.7574 Aortic elongation 0.6506 0.6792 0.7439 0.7695 0.5561 0.7303 0.3615 0.7630 0.7506 0.7613 Laminar atelectasis † 0.5544 0.5013 0.8038 0.8171 0.3144 0.7209 0.3527 0.7744 0.8374 0.8425 Costophrenic angle blunting † 0.6936 0.7096 0.7802 0.7988 0.5838 0.7758 0.5965 0.7684 0.7866 0.7928 Scoliosis 0.6448 0.6539 0.7253 0.7594 0.5085 0.6290 0.1761 0.7355 0.7900 0.7692 Chronic changes 0.6003 0.6168 0.6457 0.6579 0.6129 0.6549 0.3580 0.6858 0.6883 0.6712 Pneumonia 0.7603 0.6334 0.7875 0.7892 0.5768 0.7626 0.5659 0.7605 0.7628 0.7654 Callus rib fracture* 0.4822 0.4698 0.8172 0.7015 0.4378 0.6113 0.4273 0.7746 0.8047 0.7395 Sternotomy* 0.9319 0.9377 0.9922 0.9937 0.8969 0.9764 0.9405 0.9869 0.9908 0.9921 Unchanged 0.4610 0.4743 0.5004 0.5919 0.6113 0.5004 0.5003 0.5601 0.4853 0.6101 Vascular hilar enlargement † 0.5533 0.5131 0.5366 0.5973 0.5647 0.5851 0.4326 0.6375 0.6523 0.6996 Bronchiectasis* 0.5564 0.6251 0.7413 0.7259 0.6800 0.6408 0.7100 0.7291 0.7152 0.7320 Fibrotic band* 0.6948 0.5159 0.7974 0.8444 0.6823 0.5696 0.7795 0.7605 0.8332 0.7573 Cardiomegaly 0.7785 0.7440 0.8192 0.8276 0.6277 0.7497 0.5137 0.8046 0.8134 0.8160 Nodule † 0.6199 0.6157 0.6903 0.7597 0.5649 0.5750 0.1375 0.7033 0.7415 0.6836 79 S6.5 Youden’s J Table S16: Per-label Youden’s J Score comparison across baseline, contrastive, and autoregressive models on the PadChest dataset. Tail abnormalities are marked with an asterisk (*), and medium-prevalence abnormalities are marked with a dagger ( † ). For each row, the best value is shown in bold and the second-best value is underlined. Abnormality EVA-Base RAD-DINO ARK CheXFound BiomedCLIP MedKLIP BioViL-T Med-CLIP Med-AR-2B (ours) Med-AR-8B (ours) NSG tube † 0.8820 0.8852 0.9511 0.9492 0.7230 0.9327 0.8545 0.9364 0.9532 0.9497 Volume loss* 0.4308 0.3644 0.6642 0.6755 0.2302 0.5084 0.1155 0.6607 0.6893 0.6820 Kyphosis † 0.5430 0.4874 0.5858 0.6450 0.2261 0.5368 0.2125 0.6302 0.6426 0.6551 Alveolar pattern † 0.6048 0.5832 0.7004 0.6990 0.4071 0.6402 0.5269 0.6935 0.7029 0.6857 Pacemaker* 0.9016 0.8847 0.9755 0.9638 0.7051 0.9548 0.8210 0.9697 0.9838 0.9800 COPD signs 0.4819 0.4669 0.5064 0.5189 0.2352 0.4781 0.1556 0.5023 0.5297 0.5198 Infiltrates 0.3597 0.3360 0.4833 0.4742 0.1704 0.3990 0.2242 0.4421 0.4900 0.4621 Vertebral degenerative changes † 0.3843 0.3503 0.4312 0.5060 0.1901 0.4370 0.1486 0.5007 0.5380 0.5232 Apical pleural thickening* 0.4098 0.3553 0.5688 0.5943 0.1752 0.3897 0.1084 0.5976 0.6369 0.6053 Pleural effusion 0.7156 0.6582 0.7959 0.7967 0.5056 0.7655 0.5506 0.7864 0.7949 0.7865 Atelectasis* 0.5662 0.5275 0.6099 0.6133 0.3583 0.5812 0.4539 0.6085 0.6271 0.6026 Central venous catheter via jugular vein* 0.8794 0.8823 0.9503 0.9411 0.7718 0.9241 0.8840 0.9502 0.9622 0.9495 Air trapping † 0.4646 0.4569 0.4584 0.4846 0.1903 0.4589 0.2145 0.4779 0.4982 0.4750 Calcified granuloma* 0.3182 0.2929 0.5725 0.5220 0.1742 0.3190 0.1037 0.5521 0.6108 0.5888 Interstitial pattern 0.4896 0.4512 0.5883 0.5854 0.2990 0.5394 0.4080 0.5702 0.5917 0.5800 Aortic elongation 0.5135 0.4592 0.6076 0.6218 0.2449 0.5625 0.2346 0.6200 0.6359 0.6308 Laminar atelectasis † 0.3120 0.2745 0.5875 0.5521 0.1059 0.4064 0.1824 0.5259 0.5828 0.5627 Costophrenic angle blunting † 0.4373 0.3814 0.5940 0.5965 0.2229 0.5551 0.2959 0.5828 0.6061 0.5939 Scoliosis 0.3644 0.3239 0.4455 0.5446 0.1553 0.3712 0.1063 0.5118 0.5651 0.5705 Chronic changes 0.4371 0.4217 0.4660 0.4705 0.2827 0.4302 0.2316 0.4765 0.4993 0.4860 Pneumonia 0.4014 0.3835 0.5554 0.5485 0.1715 0.4723 0.2092 0.5058 0.5577 0.5274 Callus rib fracture* 0.2759 0.2529 0.5145 0.4184 0.1532 0.2770 0.1186 0.4975 0.5457 0.5136 Sternotomy* 0.8616 0.8945 0.9854 0.9788 0.6361 0.9534 0.8689 0.9774 0.9867 0.9840 Unchanged 0.2404 0.2216 0.2721 0.2678 0.1171 0.2440 0.1850 0.2925 0.3013 0.2989 Vascular hilar enlargement † 0.3120 0.2839 0.3705 0.3845 0.1945 0.3040 0.1990 0.3892 0.4187 0.4081 Bronchiectasis* 0.3689 0.3468 0.5345 0.5146 0.2217 0.4295 0.2145 0.5103 0.5173 0.4984 Fibrotic band* 0.3415 0.3029 0.5501 0.5343 0.2211 0.3606 0.0847 0.5132 0.5574 0.5512 Cardiomegaly 0.6155 0.5810 0.6941 0.6969 0.2913 0.6310 0.3599 0.6811 0.7020 0.6858 Nodule † 0.2543 0.2527 0.4409 0.4005 0.1308 0.2830 0.0949 0.4113 0.4392 0.4123 80