Paper deep dive
Echo2ECG: Enhancing ECG Representations with Cardiac Morphology from Multi-View Echos
Michelle Espranita Liman, ĂzgĂŒn Turgut, Alexander MĂŒller, Eimo Martens, Daniel Rueckert, Philip MĂŒller
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 97%
Last extracted: 3/13/2026, 12:51:13 AM
Summary
Echo2ECG is a multimodal self-supervised learning framework that aligns global ECG signals with multi-view echocardiography (Echo) studies to learn structure-aware ECG representations. By resolving the representational mismatch of single-view alignment and eliminating the need for paired text reports, Echo2ECG acts as a lightweight, robust feature extractor for structural cardiac phenotype classification and cross-modal retrieval.
Entities (6)
Relation Signals (4)
Echo2ECG â alignmodalities â ECG
confidence 100% · we propose Echo2ECG, a multimodal self-supervised learning framework that enriches ECG representations with the heartâs morphological structure captured in multi-view Echos.
Echo2ECG â usesencoder â OTiS
confidence 100% · The ECG encoder is a 12-layer tiny transformer [20], initialized from OTiS [19]
Echo2ECG â usesencoder â EchoPrime
confidence 100% · The Echo encoder is a MViTv2 [10], initialized from EchoPrime [21]
Echo2ECG â predicts â LVEF
confidence 95% · This enables the prediction of Echo-derived structural cardiac phenotypes, such as left ventricular ejection fraction (LVEF)
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Electrocardiography (ECG) is a low-cost, widely used modality for diagnosing electrical abnormalities like atrial fibrillation by capturing the heart's electrical activity. However, it cannot directly measure cardiac morphological phenotypes, such as left ventricular ejection fraction (LVEF), which typically require echocardiography (Echo). Predicting these phenotypes from ECG would enable early, accessible health screening. Existing self-supervised methods suffer from a representational mismatch by aligning ECGs to single-view Echos, which only capture local, spatially restricted anatomical snapshots. To address this, we propose Echo2ECG, a multimodal self-supervised learning framework that enriches ECG representations with the heart's morphological structure captured in multi-view Echos. We evaluate Echo2ECG as an ECG feature extractor on two clinically relevant tasks that fundamentally require morphological information: (1) classification of structural cardiac phenotypes across three datasets, and (2) retrieval of Echo studies with similar morphological characteristics using ECG queries. Our extracted ECG representations consistently outperform those of state-of-the-art unimodal and multimodal baselines across both tasks, despite being 18x smaller than the largest baseline. These results demonstrate that Echo2ECG is a robust, powerful ECG feature extractor. Our code is accessible at this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2603.08505v1
- Canonical: https://arxiv.org/abs/2603.08505v1
Trouble viewing inline? Open PDF directly â
Full Text
27,783 characters extracted from source content.
Expand or collapse full text
11institutetext: Chair for AI in Healthcare and Medicine, Technical University of Munich (TUM) and TUM University Hospital, Germany 11email: michelle.liman,oezguen.turgut,daniel.rueckert,philip.j.mueller@tum.de 22institutetext: Department of Cardiology, TUM University Hospital, Germany 22email: Alexander.Mueller,Eimo.Martens@mri.tum.de 33institutetext: Department of Computing, Imperial College London, UK 44institutetext: Munich Center for Machine Learning (MCML), Germany Echo2ECG: Enhancing ECG Representations with Cardiac Morphology from Multi-View Echos Michelle Espranita Liman ĂzgĂŒn Turgut Alexander MĂŒller Eimo Martens Daniel Rueckert Philip MĂŒller Abstract Electrocardiography (ECG) is a low-cost, widely used modality for diagnosing electrical abnormalities like atrial fibrillation by capturing the heartâs electrical activity. However, it cannot directly measure cardiac morphological phenotypes, such as left ventricular ejection fraction (LVEF), which typically require echocardiography (Echo). Predicting these phenotypes from ECG would enable early, accessible health screening. Existing self-supervised methods suffer from a representational mismatch by aligning ECGs to single-view Echos, which only capture local, spatially restricted anatomical snapshots. To address this, we propose Echo2ECG, a multimodal self-supervised learning framework that enriches ECG representations with the heartâs morphological structure captured in multi-view Echos. We evaluate Echo2ECG as an ECG feature extractor on two clinically relevant tasks that fundamentally require morphological information: (1) classification of structural cardiac phenotypes across three datasets, and (2) retrieval of Echo studies with similar morphological characteristics using ECG queries. Our extracted ECG representations consistently outperform those of state-of-the-art unimodal and multimodal baselines across both tasks, despite being 18Ă smaller than the largest baseline. These results demonstrate that Echo2ECG is a robust, powerful ECG feature extractor. Our code is accessible at: https://github.com/michelleespranita/Echo2ECG. 1 Introduction Electrocardiography (ECG) is a cost-effective and ubiquitous tool for cardiac assessment. While it is highly effective at diagnosing electrical heart abnormalities, such as atrial fibrillation [1], it cannot directly quantify cardiac morphological phenotypes like left ventricular ejection fraction (LVEF). Traditionally, these structural measurements require echocardiography (Echo), a modality often limited by cost, medical expertise, and accessibility. Recently, AI-based ECG models have demonstrated the ability to detect morphological phenotypes such as left-sided valvular heart disease [4] and hypertrophic cardiomyopathy [8], enabling early and accessible health screening. Current methods for transferring morphological information from Echo to ECG primarily rely on supervised training using Echo-derived labels [4, 2, 8, 13]. However, they fail to leverage large-scale unlabeled data through self-supervised learning, which is critical for building robust ECG representations that could be used for other prediction tasks. Recent multimodal self-supervised methods, most notably EchoingECG [5], attempt to bridge this gap by pre-training on ECG, Echo, and text-based discharge summaries. By aligning ECGs to single-view Echos (e.g., apical four-chamber/A4C), the approach attempts to map global electrical signals to local, spatially restricted anatomical snapshots. This creates a representational mismatch: a single-view Echo captures only a subset of the heartâs structure, whereas the ECG reflects the heartâs global electrical activity. Furthermore, by requiring paired text reports, the modelâs scalability is restricted to datasets where clinical reports are available. To address these limitations, we propose Echo2ECG, a multimodal self-supervised learning framework that enriches ECG representations with morphological information from multi-view Echo studies. By leveraging powerful unimodal encoders pre-trained on large datasets, our method enables distillation of a complete picture of the heartâs morphological structure into ECG representations without depending on paired text data. The resulting ECG encoder serves as a powerful ECG feature extractor for downstream tasks requiring morphological information. Our contributions are as follows: âą We propose Echo2ECG, a multimodal self-supervised learning framework that aligns global ECGs to multi-view Echo studies to learn rich, structure-aware ECG representations. âą We evaluate Echo2ECG as an ECG feature extractor for the classification of structural cardiac phenotypes across three datasets. Our extracted ECG features consistently outperform those of state-of-the-art unimodal and multimodal baselines, despite our model being 18Ă smaller than the largest baseline. This demonstrates that morphological information can be effectively distilled into a lightweight yet powerful ECG feature extractor. âą We demonstrate the alignment of our learned representation space by successfully retrieving Echo studies with similar morphological characteristics using only ECG queries. 2 Related Work The increasing accessibility of large-scale multimodal datasets, such as MIMIC-IV [7] and the UK Biobank [17], has enabled research into cross-modal knowledge transfer in the medical domain. Most multimodal learning frameworks are inspired by CLIP [14], which aligns modalities by pulling paired representations together and pushing unpaired ones apart in a joint embedding space. In the cardiac domain, MMCL [18] and PTACL [16] use this approach to distill cardiac morphological information from Cardiac MRI (CMR) into ECG. Most similar to our work, EchoingECG [5] aligns ECGs with single-view Echos and discharge reports. These studies consistently demonstrate that aligning ECGs with "gold standard" imaging modalities significantly improves the prediction of structural cardiac phenotypes solely based on ECG. Concurrently, large-scale ECG foundation models such as xECG [11] and ECGFounder [9], and time-series foundation models such as OTiS [19], have emerged, utilizing self-supervised learning on massive ECG datasets like MIMIC-IV-ECG [6] and CODE [15]. We argue that while these unimodal models are powerful, their performance on structural cardiac phenotype prediction can be further enhanced through multimodal alignment. To this end, we initialize the ECG encoder in our framework with OTiS [19]. Our work diverges from existing literature in two key ways. First, we align ECGs with complete multi-view Echo studies, solving the representational mismatch that occurs when aligning ECGs with individual Echo views. Secondly, unlike EchoingECG, we focus strictly on ECG-Echo relationship by omitting the usage of text reports during pre-training. To our knowledge, this is the first study to investigate the direct alignment of ECGs with full multi-view Echo studies. 3 Method Figure 1: Overview of Echo2ECG. To address the representational mismatch in prior methods aligning ECGs with single-view Echos, our framework uses contrastive learning to align ECGs with multi-view Echo studies. This multi-view alignment distills comprehensive cardiac morphological information from Echos into ECG representations. We aggregate view-level Echo embeddings from a frozen, powerful Echo encoder into a study-level Echo embedding. During pre-training, we optimize only the ECG encoder, the Echo view aggregator, and modality-specific projection layers. The ECG encoder is then used to extract ECG representations for subsequent downstream tasks. The goal of our proposed method is to transfer rich cardiac morphological information from echocardiography into ECG representations. This enables the prediction of Echo-derived structural cardiac phenotypes, such as left ventricular ejection fraction (LVEF), using the widely accessible and low-cost ECG modality alone. As shown in Figure 1, our framework comprises four main components: (1) an ECG encoder that maps an ECG signal to an embedding; (2) an Echo encoder that maps each Echo view to an embedding; (3) an Echo view aggregation module that combines multiple view-level embeddings from the same study into a single study-level representation; and (4) modality-specific projection layers. 3.1 ECG and Echo Encoders We initialize the ECG and Echo encoders using strong, pre-trained unimodal models to leverage large-scale prior knowledge. The ECG encoder is a 12-layer tiny transformer [20], initialized from OTiS [19], which was pre-trained on a multi-domain time-series corpus including MIMIC-IV-ECG [6]. Each ECG is represented by the mean token embedding. The Echo encoder is a MViTv2 [10], initialized from EchoPrime [21], which was trained via multimodal contrastive learning between Echo views and clinical reports from an in-house Cedars-Sinai dataset. Each Echo view is represented by the [CLS] token embedding. The ECG encoder, the Echo view aggregator, and the projection layers are fully trainable, while the Echo encoder is kept frozen. Overall, the framework has 12.5M trainable parameters and is trained for 50 epochs. Hyperparameter details are available in the code. 3.2 Echo View Aggregation A single Echo view captures only part of the heart, e.g., the A4C view captures the four chambers of the heart but not the aortic valve, which is visible in the PSAX view. In clinical practice, cardiologists do not base diagnoses on a single view; instead, they integrate information across multiple views from a complete Echo study to obtain a comprehensive cardiac assessment. Similarly, by distilling information from the full Echo study rather than isolated views into an ECG, we ensure that the resulting ECG features are informed by the entire heart. To achieve this, each Echo view is first encoded independently by the Echo encoder to produce view-level embeddings. These embeddings are projected from 512- to 1024-dim and aggregated via attention pooling into a single study-level embedding. This allows the model to weigh views by relevance while preserving information from the entire study. 3.3 Multimodal Contrastive Pre-Training To transfer information from multi-view Echo studies to ECGs, we align the embeddings from both modalities within a shared 512-dim latent space using the CLIP[14] contrastive objective. After modality-specific projection layers, the ECG and Echo study embeddings are pulled together for matched pairs and pushed apart for unmatched pairs within a batch. For a batch of size B, the ECGâ contrastive loss is defined as: âECGâEcho=â1Bââi=1BlogâĄexpâ(cosâ(zecgi,zechoi)/Ï)âj=1Bexpâ(cosâ(zecgi,zechoj)/Ï)L_ECG =- 1B _i=1^B exp(cos(z_ecg^i,z_echo^i)/Ï) _j=1^Bexp(cos(z_ecg^i,z_echo^j)/Ï) (1) where Ï denotes the temperature parameter, and zecgz_ecg and zechoz_echo are the ECG and Echo study embeddings after the projection layers respectively. An analogous Echoâ loss is computed by swapping the modalities. The final training objective is a weighted combination of the two directional losses. 4 Experimental Setup 4.1 Dataset and Pre-processing We pre-train using paired ECG-Echo studies collected at our private institution. An ECG and an Echo study of a patient are a valid pair if they were acquired within 7 days. We split the data at the patient level (Table 1). Table 1: Internal pre-training dataset details. Age is presented as Mean ± SD. Split #Pairs #Patients #ECGs #Echo Studies #Echo Views Age Sex (%Male) Train 11,859 6,836 11,516 8,712 64,357 65.6± 16.1 63.1% Val 1,248 565 1,213 934 7,185 65.3± 16.7 64.7% Test 1,358 566 1,290 924 7,091 64.8± 16.6 64.1% ECGs (12-lead, 10-second, 500 Hz) are processed following OTiS [19], including baseline drift removal using asymmetric least-squares smoothing and lead-group normalization. We use T=1008T=1008 timesteps and apply the same augmentations as in OTiS. Echo studies include up to 128 views. Following EchoPrime [21], we extract the Echo regions, remove annotations, and downsample the spatial resolution to 224. We sample 16 frames with stride 2, and zero-pad shorter clips. Each Echo view is normalized using EchoPrime train set statistics without augmentation. 4.2 Baselines We compare our ECG model against two categories of baselines: (1) unimodal ECG models, such as OTiS (7.1M params) [19], xECG (57.0M) [11], and ECGFounder (30.7M) [9]; and (2) multimodal ECG models, such as EchoingECG (126.6M) [5] and PTACL (19.3M) [16]. To ensure fair comparison, we follow the data pre-processing and training protocols in the original papers or codebases. 4.3 Evaluation Strategy 4.3.1 Structural Cardiac Phenotype Prediction We assess the quality of the extracted ECG representations on two downstream tasks which require morphological information: (1) left ventricular ejection fraction (LVEF) classification (reduced â€40%†40\%, mildly reduced 40â50%40-50\%, and normal â„50%â„ 50\% [12]) and (2) structural heart disease (SHD) classification (absence vs. presence of moderate/greater disease [3]). For both tasks, we train a k-nearest neighbors classifier (kNN) with k=5k=5 on the ECG features. As kNN performance is deterministic, we report AUROC with 95% CI via 1,000 bootstrap resamples of the test set. Downstream dataset details are summarized in Table 2. To assess generalization, models are evaluated across three cohorts on LVEF classification: our internal dataset, EchoNext [3], and UK Biobank (UKB) [17]. For SHD classification on EchoNext, we utilize the provided splits and evaluate the strength of the ECG embeddings by scaling training data from 0.1% to 100%. Table 2: Downstream datasets and label distributions. LVEF: 0/1/2 = reduced/mildly reduced/normal; SHD: 0/1 = absence/presence. These tasks assess how well the ECG representations capture morphological information. Task Dataset Train (N) Val (N) Test (N) LVEF clf (0/1/2) Internal 294 (33/33/33%) 564 (11/16/73%) 562 (14/15/71%) EchoNext 300 (33/33/33%) 4,112 (17/7/76%) 4,827 (16/7/77%) UKB 255 (33/33/33%) 2,834 (1/4/95%) 2,827 (1/4/95%) SHD clf (0/1) EchoNext 72,475 (48/52%) 4,626 (57/43%) 5,442 (57/43%) 4.3.2 Retrieval of Echo Studies with Similar Phenotypes using ECG Queries We assess the learned joint embedding space via phenotype-aware cross-modal retrieval. Given a query ECG, Echo study embeddings are ranked by cosine similarity, and the top-k Echo studies are retrieved. A retrieved Echo study is a match if its phenotype value lies within ±0.5âϱ 0.5Ï of the query ECGâs phenotype, where Ï denotes the standard deviation of that phenotype in the test set. For models without an Echo view aggregator, the Echo study embedding is calculated by averaging the embeddings of the Echo views belonging to the study. We evaluate four cardiac phenotypes: LV end-diastolic and end-systolic volumes (EDV, ESV), stroke volume (SV), and ejection fraction (EF); and report Precision@1 (i.e., k=1k=1) and mean rank (MnR) across the test set. 5 Results and Discussion 5.1 Echo2ECG Extracts Powerful Out-of-the-Box Features Table 3: Best and second-best AUROC [95% CI] (â ) for ECG-based LVEF classification. Echo2ECG provides high-quality, out-of-the box ECG features, achieving the highest AUROC on our internal dataset and on EchoNext. On UKB, it ranks second behind PTACL, which was pre-trained on UKB by pairing ECG with CMRâa superior modality to Echo for LVEF quantification. Paired Modality Method Internal External Internal EchoNext UKB - OTiS (7.1M) 0.688 [0.650, 0.727] 0.697 [0.680, 0.713] 0.637 [0.574, 0.694] - xECG (57.0M) 0.732 [0.696, 0.769] 0.695 [0.677, 0.711] 0.531 [0.474, 0.591] - ECGFounder (30.7M) 0.746 [0.710, 0.781] 0.686 [0.670, 0.702] 0.653 [0.592, 0.708] Text+Echo EchoingECG (126.6M) 0.740 [0.703, 0.776] 0.676 [0.661, 0.692] 0.604 [0.541, 0.662] CMR PTACL (19.3M) 0.723 [0.686, 0.759] 0.704 [0.689, 0.719] 0.729 [0.677, 0.770] Echo Echo2ECG (7.1M) 0.785 [0.750, 0.819] 0.723 [0.708, 0.738] 0.692 [0.642, 0.736] To evaluate diagnostic utility, we assess Echo2ECG on (i) LVEF (Table 3) and (i) SHD classification (Figure 2) against unimodal and multimodal baselines. Across both tasks, Echo2ECG consistently achieves superior performance. For LVEF classification, a kNN classifier trained on Echo2ECGâs ECG features achieves the highest AUROC on our internal dataset and on EchoNext, improving AUROC by 5.2% and 2.7% respectively over the second-best model on each dataset. On UKB, our model ranks second behind PTACL, which was pre-trained on UKB by pairing ECG with CMRâa superior modality to Echo for LVEF quantification. For SHD classification, a kNN classifier trained on just 1% of the ECG features extracted by Echo2ECG outperforms nearly all other models trained on the full 100% dataset (Figure 2(a)), showing that Echo2ECG provides robust ECG features that work well under low-data regimes. Remarkably, a kNN trained on just 0.1% of the ECG features extracted by Echo2ECG outperforms all competing models trained on the same amount of dataâincluding EchoingECGâwhile Echo2ECG remains 18Ă smaller (Figure 2(b)). Figure 2: AUROC (â ) for ECG-based SHD classification. (a) A kNN classifier trained on just 1% of data using ECG representations extracted by Echo2ECG outperforms most models trained on 100% of the data, meaning that Echo2ECG provides robust ECG features that work well under low-data regimes. (b) Echo2ECG outperforms EchoingECG at 0.1%0.1\% training data, despite being 18Ă smaller. 5.2 Multi-View Alignment Enables Accurate Morphological Cross-Modal Retrieval In phenotype-aware ECG-to-Echo retrieval (Table 4), Echo2ECG achieves the highest Prec@1 and lowest MnR. Our study-level, multi-view alignment outperforms single-view alignment, highlighting the importance of aligning global electrical signals with comprehensive anatomy to encode morphological structure. Interestingly, EchoingECG performs worse than random retrieval, likely because its contrastive objective prioritizes ECG-Text over ECG-Echo alignment [5], failing to directly link the heartâs electrical activity with its morphological structure. Table 4: Cross-modal retrieval performance for structural cardiac phenotypes. Multi-view alignment allows Echo2ECG to accurately retrieve Echo studies from ECG queries. Method LVEDV LVESV LVSV LVEF P@1â MnRâ P@1â MnRâ P@1â MnRâ P@1â MnRâ Random retrieval 0.296 5.915 0.366 6.081 0.323 4.570 0.295 7.230 EchoingECG 0.241 10.111 0.337 10.998 0.325 5.320 0.232 13.035 Echo2ECG (single-view Echo) 0.332 5.376 0.517 4.366 0.334 5.125 0.415 5.346 Echo2ECG (multi-view Echo) 0.387 4.358 0.517 3.841 0.379 4.193 0.404 5.101 5.3 Learnable Pooling of Multi-View Echo Benefits Alignment Our ablation study (Table 5) shows that multi-view alignment consistently outperforms single-view alignment, with attention pooling proving most effective at distilling key morphological features into ECG representations. Table 5: Ablation of Echo2ECG components. Mean AUROC [95% CI] (â ) is reported. Multi-view alignment and attention pooling improve the quality of ECG features. Alignment View Agg. LVEF clf (Internal) SHD clf (EchoNext 1%) ECG â single-view Echo - 0.747 0.743 ECG â multi-view Echo mean pool 0.775 0.752 ECG â multi-view Echo CLS pool 0.756 0.754 ECG â multi-view Echo attention pool 0.785 0.757 6 Conclusion We present Echo2ECG, a self-supervised framework that enriches ECG representations with cardiac morphology from multi-view Echo studies. Particularly, we resolve the representational mismatch of prior single-view methods and eliminate reliance on paired text data. Across three datasets, our lightweight ECG model serves as a powerful feature extractor, outperforming state-of-the-art baselines in structural phenotyping and morphology-aware cross-modal Echo retrieval. Remarkably, it achieves this while being 18Ă smaller than the largest baseline. Our work has several limitations. First, we do not account for beat- or phase-level synchronization between ECGs and Echos, which may limit the modelâs ability to capture temporal relationships between electrical activity in ECGs and mechanical activity in Echos. Second, although we do not rely on text reports, our method requires paired ECGâEcho studies, which may be scarce at scale. Acknowledgments This research has been conducted using the UK Biobank Resource under Application Number 87802. References [1] L. C.C. Brant, A. H. Ribeiro, O. B. Eromosele, M. M. Pinto-Filho, S. M. Barreto, B. B. Duncan, M. G. Larson, E. J. Benjamin, A. L.P. Ribeiro, and H. Lin (2025) Prediction of atrial fibrillation from the ecg in the community using deep learning: a multinational study. 18 (10), p. e013734. External Links: Document, Link, https://w.ahajournals.org/doi/pdf/10.1161/CIRCEP.125.013734 Cited by: §1. [2] M. Cohen-Shelly, Z. I. Attia, P. A. Friedman, S. Ito, B. A. Essayagh, W. Ko, D. H. Murphree, H. I. Michelena, M. Enriquez-Sarano, R. E. Carter, P. W. Johnson, P. A. Noseworthy, F. Lopez-Jimenez, and J. K. Oh (2021-03) Electrocardiogram screening for aortic valve stenosis using artificial intelligence. 42 (30), p. 2885â2896. External Links: ISSN 0195-668X, Document, Link, https://academic.oup.com/eurheartj/article-pdf/42/30/2885/46621476/ehab153.pdf Cited by: §1. [3] P. Elias and J. Finer (2025-09) EchoNext: A Dataset for Detecting Echocardiogram-Confirmed Structural Heart Disease from ECGs. PhysioNet. Note: Version 1.1.0 External Links: Document, Link Cited by: §4.3.1, §4.3.1. [4] P. Elias, T. J. Poterucha, V. Rajaram, et al. (2022) Deep learning electrocardiographic analysis for detection of left-sided valvular heart disease. 80 (6), p. 613â626. External Links: Document, Link, https://w.jacc.org/doi/pdf/10.1016/j.jacc.2022.05.029 Cited by: §1, §1. [5] Y. Gao, S. Kim, and C. McIntosh (2025-09) EchoingECG: An Electrocardiogram Cross-Modal Model for Echocardiogram Tasks . In proceedings of Medical Image Computing and Computer Assisted Intervention â MICCAI 2025, Vol. LNCS 15964. Cited by: §1, §2, §4.2, §5.2. [6] B. Gow, T. Pollard, L. A. Nathanson, A. Johnson, B. Moody, C. Fernandes, N. Greenbaum, J. W. Waks, P. Eslami, T. Carbonati, A. Chaudhari, E. Herbst, D. Moukheiber, S. Berkowitz, R. Mark, and S. Horng (2023-09) MIMIC-IV-ECG: Diagnostic Electrocardiogram Matched Subset. PhysioNet. Note: Version 1.0 External Links: Document, Link Cited by: §2, §3.1. [7] A. Johnson, L. Bulgarelli, T. Pollard, B. Gow, B. Moody, S. Horng, L. A. Celi, and R. Mark (2024-10) MIMIC-IV. PhysioNet. Note: Version 3.1 External Links: Document, Link Cited by: §2. [8] W. Ko, K. C. Siontis, Z. I. Attia, R. E. Carter, S. Kapa, S. R. Ommen, S. J. Demuth, M. J. Ackerman, B. J. Gersh, A. M. Arruda-Olson, J. B. Geske, S. J. Asirvatham, F. Lopez-Jimenez, R. A. Nishimura, P. A. Friedman, and P. A. Noseworthy (2020) Detection of hypertrophic cardiomyopathy using a convolutional neural network-enabled electrocardiogram. 75 (7), p. 722â733. External Links: Document, Link, https://w.jacc.org/doi/pdf/10.1016/j.jacc.2019.12.030 Cited by: §1, §1. [9] J. Li, A. D. Aguirre, V. M. Junior, J. Jin, C. Liu, L. Zhong, C. Sun, G. Clifford, M. Brandon Westover, and S. Hong (2025) An electrocardiogram foundation model built on over 10 million recordings. 2 (7), p. AIoa2401033. Cited by: §2, §4.2. [10] Y. Li, C. Wu, H. Fan, K. Mangalam, B. Xiong, J. Malik, and C. Feichtenhofer (2022) MViTv2: improved multiscale vision transformers for classification and detection. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Vol. , p. 4794â4804. External Links: Document Cited by: §3.1. [11] R. Lunelli, A. Nicolson, S. M. Pröl, S. J. Reinstadler, A. Bauer, and C. Dlaska (2025) BenchECG and xecg: a benchmark and baseline for ecg foundation models. External Links: 2509.10151, Link Cited by: §2, §4.2. [12] T. A. McDonagh, M. Metra, M. Adamo, et al. (2021-08) 2021 esc guidelines for the diagnosis and treatment of acute and chronic heart failure: developed by the task force for the diagnosis and treatment of acute and chronic heart failure of the european society of cardiology (esc) with the special contribution of the heart failure association (hfa) of the esc. European Heart Journal 42 (36), p. 3599â3726. External Links: ISSN 0195-668X, Document, Link, https://academic.oup.com/eurheartj/article-pdf/42/36/3599/40594787/ehab368.pdf Cited by: §4.3.1. [13] T. J. Poterucha, L. Jing, R. P. Ricart, et al. (2025-08-01) Detecting structural heart disease from electrocardiograms using ai. Nature 644 (8075), p. 221â230. External Links: ISSN 1476-4687, Document, Link Cited by: §1. [14] A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever (2021-18â24 Jul) Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, M. Meila and T. Zhang (Eds.), Proceedings of Machine Learning Research, Vol. 139, p. 8748â8763. External Links: Link Cited by: §2, §3.3. [15] A. L. P. Ribeiro, G. M. M. PaixĂŁo, P. R. Gomes, M. H. Ribeiro, A. H. Ribeiro, J. A. Canazart, D. M. Oliveira, M. P. Ferreira, E. M. Lima, J. L. d. Moraes, N. Castro, L. B. Ribeiro, and P. W. Macfarlane (2019-11) Tele-electrocardiography and bigdata: the CODE (clinical outcomes in digital electrocardiography) study. 57S, p. S75âS78 (en). Cited by: §2. [16] A. Selivanov, P. MĂŒller, Ă. Turgut, N. Stolt-AnsĂł, and D. Rueckert (2025-09) Global and Local Contrastive Learning for Joint Representations from Cardiac MRI and ECG . In proceedings of Medical Image Computing and Computer Assisted Intervention â MICCAI 2025, Vol. LNCS 15960. Cited by: §2, §4.2. [17] C. Sudlow, J. Gallacher, N. Allen, V. Beral, P. Burton, J. Danesh, P. Downey, P. Elliott, J. Green, M. Landray, B. Liu, P. Matthews, G. Ong, J. Pell, A. Silman, A. Young, T. Sprosen, T. Peakman, and R. Collins (2015-03) UK biobank: an open access resource for identifying the causes of a wide range of complex diseases of middle and old age. PLoS Med. 12 (3), p. e1001779 (en). Cited by: §2, §4.3.1. [18] Ă. Turgut, P. MĂŒller, P. Hager, S. Shit, S. Starck, M. J. Menten, E. Martens, and D. Rueckert (2025) Unlocking the diagnostic potential of electrocardiograms through information transfer from cardiac magnetic resonance imaging. Medical Image AnalysisNEJM AIJ. Electrocardiol.Circulation: Arrhythmia and ElectrophysiologyJACCEuropean Heart JournalJACC 101, p. 103451. External Links: ISSN 1361-8415, Document, Link Cited by: §2. [19] Ă. Turgut, P. MĂŒller, M. J. Menten, and D. Rueckert (2025) Towards generalisable time series understanding across domains. External Links: 2410.07299, Link Cited by: §2, §3.1, §4.1, §4.2. [20] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ć. Kaiser, and I. Polosukhin (2017) Attention is all you need. In Advances in Neural Information Processing Systems, I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett (Eds.), Vol. 30, p. . External Links: Link Cited by: §3.1. [21] M. Vukadinovic, I. Chiu, X. Tang, N. Yuan, T. Chen, P. Cheng, D. Li, S. Cheng, B. He, and D. Ouyang (2025-11-11) Comprehensive echocardiogram evaluation with view primed vision language ai. Nature. External Links: ISSN 1476-4687, Document, Link Cited by: §3.1, §4.1.