Paper deep dive

D2MDT: Department-aware Multidisciplinary Team Consultation with Deliberation for Efficient Clinical Prediction

Yongqi Liang, Qidong Liu, Chunze Yang, Lei Wu, Jiusong Ge, Ni Zhang, Chen Li

Year: 2026Venue: arXiv preprintArea: cs.MAType: PreprintEmbeddings: 68

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 96%

Last extracted: 7/9/2026, 1:10:21 AM

Summary

D2MDT is a department-aware multi-agent framework designed for efficient clinical prediction using electronic health records (EHRs). It addresses limitations in existing multi-agent systems, such as weak evidence differentiation and redundant multi-round interactions, by assigning patient-specific department perspectives to doctor agents and implementing residual deliberation to update only unresolved consensus. The framework fuses refined consensus reports with structured EHR representations to enhance mortality prediction performance and consultation efficiency on MIMIC-III and MIMIC-IV datasets.

Entities (13)

D2MDT · framework · 100%MIMIC-III · dataset · 99%MIMIC-IV · dataset · 99%Electronic Health Records · data-type · 98%Mortality Prediction · task · 98%Residual Deliberation · mechanism · 97%Department Agents · agent-type · 96%Qwen3-8b · model · 96%AdaCare · baseline · 95%GatorTron-base · model · 95%MedCPT · model · 95%Multi-agent systems · technology · 95%SHAP · technique · 94%

Relation Signals (12)

D2MDT → evaluateson → MIMIC-IV

confidence 99% · We evaluate D2MDT on two widely used ICU benchmark datasets, MIMIC-III (Johnson et al., 2016; Harutyunyan et al., 2019) and MIMIC-IV (Johnson et al., 2023).

D2MDT → evaluateson → MIMIC-III

confidence 99% · We evaluate D2MDT on two widely used ICU benchmark datasets, MIMIC-III (Johnson et al., 2016; Harutyunyan et al., 2019) and MIMIC-IV

D2MDT → targets → Mortality Prediction

confidence 98% · Experiments on mortality prediction show that D2MDT improves both predictive performance and consultation efficiency.

D2MDT → uses → Electronic Health Records

confidence 98% · Electronic health records (EHRs) are central to clinical prediction... D2MDT first constructs structured EHR evidence and consultation-ready semantic evidence

D2MDT → introduces → Residual Deliberation

confidence 97% · To improve efficiency, D2MDT further introduces residual deliberation, which updates only unresolved consensus rather than replaying the full discussion history.

D2MDT → assigns → Department Agents

confidence 96% · It then assigns patient-specific department perspectives to doctor agents and retrieves complementary evidence for collaborative consultation.

D2MDT → utilizes → Qwen3-8b

confidence 96% · All consultation agents are driven by Qwen3-8B.

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Electronic health records (EHRs) are central to clinical prediction, but existing methods either rely on correlation-driven deep models or use single large language models (LLMs), making it difficult to support multidisciplinary clinical reasoning. Recent multi-agent systems (MAS) provide a promising alternative, yet current EHR-grounded MAS methods still suffer from weak evidence differentiation across agents and redundant multi-round interaction. We propose D2MDT, a Department-aware MultiDisciplinary Team Consultation with Deliberation for Efficient clinical prediction. D2MDT first constructs structured EHR evidence and consultation-ready semantic evidence for multi-agent consultation. It then assigns patient-specific department perspectives to doctor agents and retrieves complementary evidence for collaborative consultation. To improve efficiency, D2MDT further introduces residual deliberation, which updates only unresolved consensus rather than replaying the full discussion history. Finally, D2MDT fuses the refined consensus report with structured EHR representations for prediction. Experiments on mortality prediction show that D2MDT improves both predictive performance and consultation efficiency. We release the code online to ease the reproducibility of this paper.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

67,236 characters extracted from source content.

Expand or collapse full text

D 2 MDT: Department-aware Multidisciplinary Team Consultation with Deliberation for Efficient Clinical Prediction Yongqi Liang, Qidong Liu * , Chunze Yang, Lei Wu, Jiusong Ge, Ni Zhang, Chen Li * Xi’an Jiaotong University, Xi’an, China LYQi, pureeeee, 2216113083, jiusongge@stu.xjtu.edu.cn liuqidong, nizhang, cli@xjtu.edu.cn * Corresponding authors. Abstract Electronic health records (EHRs) are central to clinical prediction, but existing methods ei- ther rely on correlation-driven deep models or use single large language models (LLMs), mak- ing it difficult to support multidisciplinary clin- ical reasoning. Recent multi-agent systems (MAS) provide a promising alternative, yet current EHR-grounded MAS methods still suf- fer from weak evidence differentiation across agents and redundant multi-round interaction. We propose D 2 MDT, a Department-aware MultiDisciplinary Team Consultation with Deliberation for Efficient clinical prediction. D 2 MDT first constructs structured EHR ev- idence and consultation-ready semantic evi- dence for multi-agent consultation. It then as- signs patient-specific department perspectives to doctor agents and retrieves complementary evidence for collaborative consultation. To im- prove efficiency, D 2 MDT further introduces residual deliberation, which updates only unre- solved consensus rather than replaying the full discussion history. Finally, D 2 MDT fuses the refined consensus report with structured EHR representations for prediction. Experiments on mortality prediction show that D 2 MDT im- proves both predictive performance and consul- tation efficiency. We release the code online to ease the reproducibility of this paper 1 . 1 Introduction Electronic health records (EHRs) have become a central foundation for data-driven clinical decision support, where patient trajectories are modeled to anticipate future risks (Zeng et al., 2025) and sup- port timely intervention (Chaturvedi et al., 2025). Existing deep learning methods generally learn pre- dictive patterns directly from EHR sequences, such as AdaCare (Ma et al., 2020) and PAI (Liao et al., 2024). However, most deep EHR models remain primarily correlation-driven, making it difficult to 1 https://github.com/GigiResearch/D2MDT explicitly capture clinical semantics and incorpo- rate medical knowledge. This limitation has mo- tivated growing interest in Large Language Mod- els (LLMs)-based EHR modeling, where language models can connect structured patient records with semantic reasoning and natural-language interpre- tation (Ren et al., 2025; Du et al., 2026). While LLMs offer potential for semantic EHR modeling, complex clinical decision-making is in- herently collaborative rather than individual (Lee et al., 2024; Hager et al., 2024). In general, difficult cases require multidisciplinary team (MDT) consul- tation, where experts with complementary knowl- edge jointly examine clinical evidence, compare di- agnostic hypotheses, and refine treatment decisions through structured discussion (Balogh, 2015; Cam- pione Russo et al., 2025; Graber et al., 2017; Wal- raven et al., 2022). This indicates the necessity of MDT in the EHR prediction task. However, MDT consultation is difficult to replicate by single LLMs. Despite LLMs’ ability to understand cases by clin- ical semantics, it still tends to compress the case into one dominant reasoning trajectory, making it difficult to preserve parallel specialty-specific as- sessments or coordinate iterative cross-perspective discussion (Hager et al., 2024; Bds, 2024). Recently, Multi-Agent Systems (MAS) have emerged as a promising approach for simulating MDT-style clinical reasoning in healthcare by in- corporating diverse clinical perspectives and inter- pretable intermediate discussions. Earlier research studies (Kim et al., 2024; Wang et al., 2024; Nori et al., 2025; Liu et al., 2026), denoted as Linguis- tic MAS, show that multi-agent deliberation can improve semantic clinical reasoning. However, these frameworks operate primarily over textual case understanding, so their reasoning remains only weakly tied to structured patient measurements and is therefore limited for rigorous EHR-based pre- diction. To address this issue, EHR-aware MAS works, such as ColaCare (Wang et al., 2025), and arXiv:2606.03543v1 [cs.MA] 2 Jun 2026 MAS Discussion & Consensus ... MultidisciplinaryTeam ... EHR Data Residual Deliberation Targeted RevisionPreliminary Consensus Prediction & Report D 2 MDT Multi-Agent Collaboration Department Assignment leaderHemodynamicOxygenationMetabolic doctor1doctor2doctor3doctor4 Prediction & Report EHR Data Weak Evidence Differentiation Redundant Multi- round Interaction x x x x x x x x x Accumulate Discussion History Figure 1: The illustration of D 2 MDT and existing MAS. ClinNoteAgents (Zhou et al., 2025a) strengthen EHR grounding and bring collaborative reasoning closer to quantitative patient-level prediction. However, there are still two challenges faced by EHR-aware MAS methods, as illustrated in Fig- ure 1. i) Weak Evidence Differentiation. Multi- agent frameworks can improve apparent diversity by assigning different roles to different agents (Shi et al., 2024; Kim et al., 2024; Tang et al., 2024; Jiang et al., 2025; Zhu et al., 2025), but these roles are usually only weakly grounded in patient- specific structured evidence, and therefore role di- versity does not necessarily translate into evidence diversity. i) Redundant Multi-round Interaction. As consultation rounds accumulate, later interac- tions repeatedly revisit resolved content, leading to substantial context redundancy and reduced rea- soning efficiency (Wang et al., 2025; Chen et al., 2025a). These two issues jointly limit both the re- liability and the scalability of multi-round clinical consultation for EHR prediction. To address the challenges mentioned above, we propose D 2 MDT, a Department-aware MultiDis- ciplinary Team Consultation with Deliberation for Efficient clinical prediction. First, D 2 MDT intro- duces Department Agents to handle the issue of weak evidence differentiation. Instead of relying on loosely specified agent roles, D 2 MDT assigns patient-specific department perspectives to doctor agents and couples them with department-aware evidence retrieval. Then, D 2 MDT reduces redun- dant multi-round interaction by carrying forward only the unresolved part of the current consensus, allowing later rounds to focus on disagreement re- finement rather than repeated discussion of the full patient context. The contributions are as follows: •Insightfully, we introduce a department-aware collaborative agent framework that explicitly sim- ulates department-level consultation for EHR- based clinical prediction. •Technically, we design a fusion mechanism that integrates Department Agent reasoning with model-agent predictions, enabling structured co- ordination between numerical EHR models and language-based clinical reasoning. •We empirically evaluate D 2 MDT on two EHR datasets and show that it achieves better predic- tive performance with higher consultation effi- ciency in multi-round settings. 2 Preliminary In this section, the problem definition of the pre- diction task is given. Specifically, we focus on predicting patient mortality outcomes. Our goal is to extract knowledge from EHR data, supple- mented by external medical knowledge (such as clinical guidelines), to enhance predictive model- ing of electronic health records. Therefore, the prediction target is expressed as: ˆy = G(x EHR , MedicalKnowledge)(1) whereˆyis the prediction outcome. For example, ˆy = 0indicates the patient is alive in the mortality prediction task.G(·,·)is the department frame- work.x EHR is the structure of the EHR dataset which is multivariate time series data with multiple features, represented asX = [x 1 , x 2 ,..., x t ] ⊤ ∈ R T×F , encompassing information acrossTvisits andFfeatures, including dynamic features (e.g., laboratory tests and vital signs). 3 Method 3.1 Overview We propose D 2 MDT, a Department-aware Mul- tidisciplinary Team Consultation with Delibera- tion for Efficient clinical prediction. As illustrated in Figure 2, to address the weak evidence differ- entiation and redundant multi-round interaction, D 2 MDT first performs Evidence Modeling to de- rive structured EHR evidence and consultation- ready semantic evidence. However, evidence views alone do not guarantee clinically meaningful spe- cialist diversity. D 2 MDT therefore conducts De- partment Collaborative Consultation, where doc- tor agents are assigned patient-specific department perspectives and access complementary external evidence to form an initial patient-level consen- sus. On top of this, D 2 MDT applies Residual Deliberation to carry forward only the unresolved part of the current consensus, so that later rounds Residual Consensus Update Residual Deliberation EHR Evidence Initial Consensus Department Collaborative Consultation Department-aware Evidence Retrieval Department Assignment EHR Data Expert Models Text Encoder Expert Embedding Report Embedding Multimodal Fusion Prediction LeaderAgent DoctorAgent3 DoctorAgent1 DoctorAgent2 Residual State Construction Final Report Semantic Evidence Consensus-Guided Multimodal Prediction Evidence Modeling Figure 2: The overview of the proposed D 2 MDT. refine disagreement through compact updates in- stead of replaying the entire interaction history. Fi- nally, D 2 MDT performs Consensus-Guided Mul- timodal Prediction by fusing the refined consen- sus report with structured EHR representations for final risk estimation. In this way, D 2 MDT turns col- laboration into an evidence-grounded and refined reasoning process for clinical prediction. 3.2 Evidence Modeling This section aims to build evidence of D 2 MDT. Since consultation should be grounded in patient- specific clinical signals rather than latent vectors alone, we construct both structured EHR Evidence for prediction and Semantic Evidence for Multi- Agent consultation from records. EHR Evidence. For a patient, letbdenote the basic profile, including static demographics and admission-related information, LetE m M m=1 de- note the set of pretrained EHR expert models, where the m-th expert E m contains a temporal en- coder and its prediction head. D 2 MDT applies all experts to the same patient trajectory and obtains: h m , ˆy m = E m (X) m = 1,...,M,(2) whereh m ∈R d m is the structured patient represen- tation produced by expertE m , andˆy m ∈ (0, 1)is its preliminary risk estimate. These expert-specific hidden representations preserve complementary temporal patterns and are later used in multimodal fusion. During fusion training, the pretrained ex- pert models are kept fixed. Semantic Evidence. Numerical embeddings alone are insufficient for language-based consultation. D 2 MDT therefore converts each expert output into an interpretable evidence card in two steps. For expertE m , we first compute feature attributions by α m = SHAP(E m , X)(Sundararajan and Najmi, 2020), where SHAP estimates feature attribution scores based on Shapley values to quantify each variable’s contribution to the expert model’s risk estimate. We retain the top-K m important variables f m,k . We then verbalize each retained variable with its name, latest value, and short local trend near timeT(e.g., persistent elevation, recent drop, or abnormal fluctuation). We denote the resulting expert-specific evidence card by: E m =(f m,k ,u m,k ,α m,k ) K m k=1 (3) whereu m,k is the textual status summary, andα m,k is the attribution score measuring how strongly this variable supports the expert prediction. D 2 MDT then forms the expert consultation context: c m = b; Recent(X); ˆy m ;E m (4) whereRecent(·)denotes a short summary of the most recent observations and local temporal changes. The resulting context summarizes the pa- tient profile, expert-estimated risk, and salient ab- normalities to expertE m . For department routing, D 2 MDT further uses the patient-level summary c route = [b; Recent(X);(ˆy m ,E m ) M m=1 ], which aggregates all expert-specific evidence cards into a unified routing view. The retained fields are pro- vided in Appendix A.1 for implementation clarity. 3.3 Department Collaborative Consultation Building on the evidence views above, D 2 MDT or- ganizes multi-agent consultation into a department- aware specialist collaboration process. This pro- cess consists of three consecutive stages: Depart- ment Assignment, which allocates specialist iden- tities to doctor agents; Department-aware Evi- dence Retrieval, which enables each specialist to access complementary external evidence under its assigned department perspective; and Initial Con- sensus, which aggregates the resulting specialist opinions into a patient-level preliminary report. Department Assignment. D 2 MDT constructs three routing cues. First, it extracts the set of abnor- mal measurementsU abn from the patient trajectory by flagging variables whose latest value deviates markedly from that patient’s recent range. Second, it aggregates the expert-level important feature sets U imp = S M m=1 f m,k into a patient-level impor- tant feature set,f m,k is the important feature set of expertE m , Third, it extracts a context-based risk signalc d i from patient-level summaryc route , which preserves coarse global cues such as pre- dicted mortality risk, ICU status, and high-risk per- centile information. For each candidate department d i ∈D, D 2 MDT computes a routing score: s d i = X f∈U abn λ abn ρ d i (f) + X f∈U imp λ imp ρ d i (f) + λ ctx c d i (5) whereρ d i (·)denotes the match score between vari- ablefand departmentd i . Therefore, the three terms correspond to abnormal-numeric evidence, important-feature evidence, and global context-risk evidence, respectively.λ abn ,λ imp , andλ ctx are their balancing weights. The assigned departments are then selected by ranking these scores: (d 1 ,...,d M ) = Top M s d | d i ∈D (6) In implementation, one hemodynamic role is kept as a global anchor, while the other doctor agents are filled by the highest-scoring department candidates. More details are given in Appendix A.2. Department-aware Evidence Retrieval. After department assignment, each doctor agent reasons over the same patient case through a department- specific clinical lens.For doctor agentD m , D 2 MDT constructs a department-aware consulta- tion context by adding a department hint, and con- structs the retrieval queryq m by appending depart- ment keywords to the patient summary. In this way, the same patient case is viewed under different specialist emphases. The exact prompt and tem- plates are listed in Appendix A.3. The doctor then retrieves complementary medical evidence by: z m = Retrieve(q m )(7) whereq m denotes the resulting department-aware query andz m denotes the corresponding retrieved evidence. This design ensures that different doctor agents differ not only in role identity but also in the supporting evidence they consult. Initial Consensus. Given the department-aware patient view and the corresponding retrieved exter- nal evidencez m , them-th doctor agent produces a structured specialist opinion: o (0) m = Doctor m (b,z m )(8) whereo (0) m contains the doctor’s risk judgment, sup- porting rationale, and cited evidence grounded in the department-specific patient view. The leader agent then aggregates all specialist opinions into a preliminary patient-level report: R (0) = Leader b,o (0) m M m=1 (9) This report serves as the initial consensus for the subsequent residual deliberation process. 3.4 Residual Deliberation Although department collaborative consultation im- proves evidence diversity, later-round interaction can still become inefficient if all agents continue ex- changing full narrative histories. D 2 MDT therefore proposes a residual deliberation process, in which each round focuses only on the unresolved part of the current consensus. Concretely, this process in- volves two steps: Residual State Construction, which separates stable consensus from remaining disagreement, and Residual Consensus Update, which guides doctor agents to revise their opinions based on the residual disagreement. Residual State Construction. After roundr, the leader reportR (r) is compressed into a settled con- sensus stateR (r) C and a residual disagreement state R (r) ∆ ,R (r) C stores stable fields such as current risk level, consensus strength, and agreed mechanisms, whileR (r) ∆ keeps only unresolved discussion points such as minority view, risk disagreement, mecha- nism mismatch, or escalation cues. Each specialist opinion is compressed into a compact doctor state: s (r) m = v (r) m ; μ (r) m ; ψ (r) m ; η (r) m ; γ (r) m m = 1,...,M. (10) wherev (r) m is the doctor’s risk vote,μ (r) m summa- rizes top mechanisms,ψ (r) m records key support features,η (r) m records counter-evidence, andγ (r) m represents confidence and escalation status. As a result, later rounds operate on compact structured summaries rather than full narrative histories. Residual Consensus Update. At roundr + 1, doctor agentD m updates its opinion by reusing its department-aware patient view and retrieved external evidence while receiving only the compact residual messages from the previous round: o (r+1) m = Doctor m s (r) m ,R (r) ∆ ,R (r) C (11) Heres (r) m ,R (r) ∆ , andR (r) C are compact residual states transmitted between rounds. Thus, later rounds do not replay earlier detailed opinions; they only pass the concise doctor card and compact leader report. The leader then revises the global report by integrating the specialist updates with the current consensus and residual disagreement state: R (r+1) = Leader(R (r) C ,R (r) ∆ ,o (r+1) m M m=1 (12) Therefore, D 2 MDT performs multi-round consul- tation as a sequence of residual corrections, where each round only refines what remains unresolved. The discussion terminates when the residual state becomes empty or sufficiently weak, or when the maximum number of rounds is reached. LetK denote the total number of executed rounds and let R ∗ = R (K) be the final consensus report. 3.5Consensus-Guided Multimodal Prediction The final consensus report is not used only for in- terpretability, it is also treated as an additional pre- dictive modality. D 2 MDT first encodes the final consultation report by a clinical text encoder: h txt = E txt (R ∗ )(13) whereE txt (·)denotes the report encoder andh txt is the resulting text representation. D 2 MDT then fuses this representation with the structured EHR representations from all expert models: h fus = φ([h 1 ;· ; h M ; h txt ])(14) where[·;·]denotes vector concatenation andφ(·)is a fusion network. The final prediction is as follows: ˆy = σ(w ⊤ h fus + b 0 )(15) wherewandb 0 are the predictor parameters, and σ(·)is the sigmoid function. For a training set with Npatients, the model is optimized by the binary cross-entropy objective: L pred =− 1 N N X i=1 (y i log ˆy i + (1− y i ) log(1− ˆy i )) (16) wherey i ∈0, 1is the ground-truth clinical out- come of patienti, andˆy i is the predicted risk. This design allows D 2 MDT to combine fine-grained temporal EHR dynamics with high-level consensus reasoning formed through department-aware and residual-efficient consultation. 3.6 Training and Inference Training. Each specialist EHR encoder is trained independently to extract patient-specific hidden rep- resentations for all samples. The fusion predictor then takes the specialist EHR representations to- gether withh txt as input and is optimized with the binary cross-entropy loss in Eq.(16). The pre- trained specialist EHR encoders are kept fixed. Inference. Given a patient’s EHR sequence, the pretrained specialist encoders first produce the hid- den representationsh m . The department collabo- rative consultation module then generates the final consensus report through multi-round reasoning, and the report is encoded ash txt . Finally, the fu- sion module combinesh m andh txt to produce the mortality predictionˆy. Due to space limitation, we leave the algorithm to Appendix A.4. 4 Experiment 4.1 Experimental Setup Datasets and Task. We evaluate D 2 MDT on two widely used ICU benchmark datasets, MIMIC- I (Johnson et al., 2016; Harutyunyan et al., 2019) and MIMIC-IV (Johnson et al., 2023). For both datasets, the task is in-hospital outcome prediction based on structured EHR records. Dataset statistics, splits, and details are deferred to Appendix B.1. Baselines.We compare D 2 MDT with three groups of baselines. The first group contains deep learning baselines, including AdaCare (Ma et al., 2020), RETAIN (Choi et al., 2016b), and Table 1: The overall results of competing baselines and D 2 MDT on MIMIC-I Outcome and MIMIC-IV Outcome. The boldface refers to the highest score and the underlineindicates the second-best result. Methods MIMIC-I OutcomeMIMIC-IV Outcome AUPRC (↑)AUROC (↑)min(+P, Se) (↑)AUPRC (↑)AUROC (↑)min(+P, Se) (↑) AdaCare48.24±3.2380.30±1.7046.45±3.3860.51±4.4785.85±2.1457.35±3.90 RETAIN50.51±4.4284.23±1.7350.88±3.7457.43±5.5587.10±2.1458.04±4.41 PAI 49.72±4.1181.76±1.9149.77±3.2354.68±5.5685.07±2.3557.65±4.47 EMERGE50.82±3.7483.49±1.7250.35±2.6258.74±4.3987.85±1.6857.27±3.38 MedGemma ZeroShot 28.64±3.2169.11±2.5831.16±2.8728.61±3.7473.31±2.8534.52±3.92 MedGemma FewShot 14.97±1.2456.28±1.4215.18±1.2611.79±1.1539.10±1.5312.43±1.21 MDAgents45.72±4.4482.31±1.8948.41±3.4253.49±5.1285.57±1.9754.42±4.17 MDTeamGPT48.31±3.3280.19±1.8844.62±2.9356.87±4.881.47±2.653.05±4.2 ColaCare50.81±5.0584.07±2.0150.52±4.1963.24±5.3988.17±2.1961.03±4.52 D 2 MDT53.72±4.6784.56±2.0754.68±3.7065.31±4.4187.39±1.9763.66±3.91 PAI (Liao et al., 2024). The second group contains LLM-driven collaborative baselines, including EMERGE (Zhu et al., 2024) and MedGemma (Sel- lergren et al., 2025). The third group contains MDT baselines, including MDAgents (Kim et al., 2024) MDTeamGPT (Chen et al., 2025a) and Co- laCare (Wang et al., 2025). Appendix B.2 gives the details of baseline settings. Evaluation Metrics. We report AUPRC (Kim and Hwang, 2022), AUROC (McDermott et al., 2024), and min(+P, Se) (Ma et al., 2022) for pre- dictive performance. To evaluate efficiency, we additionally report Average Round-1 Prompt To- ken (AR1PT) and Average Round-2 Prompt Token (AR2PT). AR2PT being smaller than AR1PT indi- cates that residual discussion is effective during the collaboration process. More details of evaluation metrics are given in the Appendix B.3. Implementation Details. All consultation agents are driven by Qwen3-8B. For retrieval-augmented consultation, we use MedCPT with the MSD cor- pus. For multimodal prediction, the final leader re- port is encoded by GatorTron-base and fused with the three expert EHR embeddings by a lightweight MLP with hidden dimension 128. Additional de- tails are provided in the Appendix B.4. 4.2 Overall Performance We show the overall performance of our D 2 MDT and competing baselines in Table 1.Overall, the proposed D 2 MDT achieves the best perfor- mance on most key metrics across the two datasets, while remaining competitive on the remaining ones, demonstrating its overall effectiveness for EHR- based outcome prediction. Then, a more detailed analysis of the results will be given. Deep Learning baselines lag behind the stronger methods overall. This is mainly because they are limited in capturing complex clinical semantics and leveraging complementary knowledge from multiple medical perspectives. As a result, their performance improvements are constrained when facing challenging outcome prediction tasks. LLM-based methods do not show clear advan- tages in this task. Although such methods intro- duce external medical knowledge or reasoning abil- ity from LLMs, they still struggle to fully adapt to structured and temporal clinical prediction settings. MDT-based baselines perform better overall, which verifies the value of multi-disciplinary col- laboration in clinical decision-making. By intro- ducing interactions among different medical roles, these methods are able to provide more comprehen- sive evidence than conventional deep learning or vanilla LLM-based approaches. However, D 2 MDT still shows stronger overall performance than these baselines, suggesting that our method can more ef- fectively transform multi-disciplinary discussions into predictive signals. The advantage of D 2 MDT indicates that structured collaboration and effec- tive aggregation of expert opinions are crucial for improving clinical outcome prediction, especially under complex and high-risk medical scenarios. 4.3 Efficiency Study We show the efficiency comparison of D 2 MDT and competing MDT-based baselines in Figure 3, using AR1PT and AR2PT defined in Evaluation Metrics. D 2 MDT achieves the best overall efficiency on both MIMIC-I Outcome and MIMIC-IV Outcome, es- pecially in the second-round consultation, which AR1PTAR2PT 0 5 10 15 Prompt Tokens (K) 7.3 14.8 8.7 15.2 8.1 7.3 MIMIC-I Outcome AR1PTAR2PT 0 5 10 7.2 12.1 9.8 10.9 9.2 6.5 MIMIC-IV Outcome MDAgentsColaCareD²MDT Figure 3: The results of efficiency study. 2345 max_round 51.0 52.5 54.0 55.5 AUPRC (%) (a) MIMIC-I AUPRC 2345 max_round 62.4 63.2 64.0 64.8 65.6 AUPRC (%) (b) MIMIC-IV AUPRC Figure 4: Sensitivity analysis of D 2 MDT with respect to max_round, measured by AUPRC on two datasets. demonstrates the effectiveness of our Residual De- liberation. Although D 2 MDT keeps competitive prompt cost in the first round, its main advantage lies in reducing prompt burden in later rounds. Compared with other MDT-based baselines, MDAgents and ColaCare usually require more prompt tokens to continue the consultation pro- cess. This is mainly because these methods tend to propagate more redundant discussion context across rounds. In contrast, D 2 MDT focuses on disagreement-aware communication and preserves only the most decision-relevant information for subsequent consultation. Therefore, D 2 MDT can maintain effective collaboration while significantly reducing unnecessary communication overhead. 4.4 Ablation Study To investigate whether each designed component contributes to D 2 MDT, we evaluate several ablated variants in Tables 2 and 3. • w/o HEM removes heterogeneous expert model- ing. The performance drops clearly, especially in AUPRC and min(+P, Se), showing that hetero- geneous experts help capture different aspects of patient status for outcome prediction. • w/o RD removes residual deliberation. This leads to higher second-round prompt cost and weaker overall results, indicating that residual informa- tion helps preserve useful evidence while reduc- ing redundant discussion. •w/o DR removes department-aware evidence re- trieval. This variant uses less targeted evidence. Its performance decrease shows that department- Table 2: Ablation study results on MIMIC-I. Model MIMIC-I Outcome AUPRC (↑) AUROC (↑) min(+P, Se) (↑) AR1PTAR2PT w/o HEM51.87±4.55 84.25±1.9450.39±3.8411271.37 8925.69 w/o RD53.36±4.31 84.43±1.9453.36±3.798523.41 11874.77 w/o DR51.77±3.73 83.22±1.9252.32±3.118559.637200.46 w/o DA52.51±4.88 84.87±1.7553.70±3.568645.767608.78 D 2 MDT53.72±4.6784.56±2.0754.68±3.708145.647344.54 Table 3: Ablation study results on MIMIC-IV. Model MIMIC-IV Outcome AUPRC (↑) AUROC (↑) min(+P, Se) (↑) AR1PT AR2PT w/o HEM57.93±5.32 86.44±2.6460.91±4.789698.24 7404.29 w/o RD65.15±5.04 87.68±2.0161.44±4.159028.31 9178.33 w/o DR62.99±4.72 87.70±2.2061.16±3.989328.46 7846.84 w/o DA63.34±5.11 87.47±2.3860.68±4.589704.42 7901.63 D 2 MDT65.31±4.4187.39±1.9763.66±3.919164.026534.88 specific retrieval helps each doctor ground the discussion in more relevant clinical context. •w/o DA removes department assignment. It ver- ifies that explicit department specialization im- proves collaboration quality. 4.5 Hyper-parameter Analysis We vary the maximum number of consultation rounds (i.e.,max_round), and report the corre- sponding results in Figure 4. Since AUPRC is the primary metric for imbalanced clinical outcome prediction, we present only the AUPRC trends in the main text for clarity, while the results of other metrics are deferred to Appendix B.6. As shown in Figure 4, the two datasets exhibit different trends. On MIMIC-I Outcome, the AUPRC increases steadily asmax_roundgrows, showing that additional consultation rounds help the doctor agents refine their judgments and reach better consensus. In contrast, On MIMIC- IV Outcome, the best AUPRC is achieved at max_round=3, and further increasing the number of rounds leads to performance degradation. This result suggests that most useful collaborative in- formation has already been exchanged in the early rounds, while excessive discussion may introduce redundant context and reduce decision quality. Considering both predictive performance and the resource cost of additional consultation rounds, we set max_round=3 for D 2 MDT. 4.6 Case Study To evaluate whether D 2 MDT produces reliable and interpretable consultation traces, we analyze a representative MIMIC-I test patient with cross- system abnormalities. The three doctor agents are In my opinion, the patient has a low risk of mortality. ... ... key_mechanisms: hemodynamic instability/perfusion compromise/ oxygenation compromise; support_summary: Medical Issues in Long-Term Renal Replacement .... .... | Document 1 (Title: Critically Ill Patient Scoring Systems) Scoring systems have been used to assess ... ...Document 2... ...Document 3... ...; answer: disagree, reason: The patient's risk prediction by the AdaCare model is 0.19, placing her in the top 30% of risk, which indicates a relatively high risk of mortality......, evidences: The absence of critical values ....... answer: agree, reason: In my initial analysis, I focused on ......, which suggested a high mortality risk. However, ......the mortality risk is not necessarily high......, evidences: Higher APACHE I scores correlate with ...... answer: agree, reason: The model's emphasis on pH, systolic blood pressure...... but The leader doctor's synthesized report provides a more balanced view...... , evidences: APACHE I combines acute ...... action: Yes, answer: In my opinion, the patient has a high risk of mortality... ..., final_risk: high; consensus_level: strong; key_mechanisms: hemodynamic instability / oxygenation compromise / metabolic stress; discussion_points: address minority disagreement evidence, reasons: Doctor 1's opinion highlights the high mortality risk... ....... ... answer: agree, reason: These factors significantly increase the patient's risk ......, evidences: RTA is a chronic metabolic ...... answer: agree, reason: The high risk is further justified by the ......., evidences: Document 1 states that ....... answer: agree, reason: These factors support the high-risk consensus ...... ., evidences: Document 1 discusses ....... In my opinion, the patient has a high risk of mortality.Doctor 1's opinion supports the high mortality risk by emphasizing the clinical manifestations of chronic RTA, such as ... ...Doctor 2's opinion reinforces the high mortality risk by highlighting the importance of addressing oxygenation compromise, particularly ... ... Doctor 3's opinion aligns with the high mortality risk by discussing the metabolic complications of chronic RTA and long-term dialysis, such as .... ... Doctor1 (hemodynamic) Doctor2 (oxygenation) Doctor3 (metabolic) Initial Consensus Final Report Leader Figure 5: Representative D 2 MDT case study. The orange and blue modules denote Doctor Agents and the Leader Agent, respectively, while gray boxes show the initial and final reports during the consultation process. assigned to critical care/hemodynamic, respira- tory/oxygenation, and metabolic/general perspec- tives, respectively, so that the same EHR trajectory is examined through different clinical roles. As illustrated in Figure 5, initial Consensus indi- cates the patient has a low mortality risk. In round1, each doctor revisits initial consensus and their own state to express agreement or disagreement. Then leader keeps the disagreement as a compact resid- ual package rather than replaying the full dialogue. In round 2, each doctor revisits only the unresolved minority evidence, and the team finally converges withminority_view: none. This example high- lights two properties of D 2 MDT. First, the decision path remains interpretable because each opinion is explicitly grounded in a department role and concrete physiological evidence. Second, the resid- ual deliberation mechanism improves efficiency by shrinking the follow-up prompt tokens. The full multi-panel qualitative is placed in Appendix B.7. 5 Related Works Retrieval- and LLM-enhanced Clinical Predic- tion. Recent studies incorporate medical knowl- edge and language models into EHR prediction. RAM-EHR (Xu et al., 2024) uses retrieved medical knowledge, CPLLM (Ben Shoham and Rappoport, 2024) adapts LLMs for disease and readmission prediction, and LLM-based EHR encoders (Hegsel- mann et al., 2025) serialize structured records into natural language. Clinical LLM systems such as EMERGE (Zhu et al., 2024) and MedGemma (Sel- lergren et al., 2025) further show the potential of LLMs for decision support. Nevertheless, most methods use LLMs as a unified reasoning or en- coding module, rather than maintaining parallel specialty-specific assessments. MDT-inspired and Agentic Clinical Reasoning. LLM-based clinical agents emulate collaborative decision-making through role specialization, de- bate, and consensus.Systems such as MDA- gents (Kim et al., 2024), MAC (Chen et al., 2025b), MAM (Zhou et al., 2025b), MDTeamGPT (Chen et al., 2025a), and ColaCare (Wang et al., 2025) demonstrate the value of complementary clinical opinions. However, their predefined roles are of- ten weakly grounded in patient-specific structured EHR evidence, and multi-round discussions may revisit resolved information. D 2 MDT addresses these limitations with department-aware evidence organization and residual deliberation. 6 Conclusion In this paper, we propose D 2 MDT, a Department- aware MultiDisciplinary Team Consultation with Deliberation for Efficient Clinical Prediction. Ex- tensive experiments on MIMIC-I and MIMIC- IV demonstrate that D 2 MDT achieves competitive predictive performance while reducing prompt to- ken consumption. In the future, we will explore stronger medical foundation models, more accurate clinical role modeling, and more adaptive consulta- tion strategies for complex healthcare scenarios. Limitations This work has several limitations. First, our evalua- tion is conducted on two ICU outcome prediction benchmarks, MIMIC-I and MIMIC-IV. Although these datasets are widely used for EHR modeling, the current results are still limited to retrospective in-hospital mortality prediction. The effectiveness of D 2 MDT on other clinical tasks, patient popula- tions, and real-world deployment scenarios remains to be further studied. Second, D 2 MDT relies on LLM-based agents and retrieved medical knowledge to conduct department-aware consultation. Therefore, its rea- soning quality can still be affected by the capa- bility of the backbone LLM, the coverage of the external knowledge corpus, and the accuracy of department routing. In addition, although residual deliberation reduces redundant discussion, it may omit useful context when the disagreement state is overly compressed. Future work will explore more adaptive routing and residual construction strate- gies, together with clinician evaluation in realistic clinical workflows. Third, as a clinical prediction framework, D 2 MDT may pose potential risks if used beyond research settings. Incorrect predictions or over- confident agent-generated rationales could mislead downstream clinical decision making, especially if the system is treated as a substitute for clinician judgment. In addition, biases inherited from retro- spective ICU datasets may lead to uneven perfor- mance across patient subgroups, and LLM-based agents may generate incomplete or unsupported reasoning when the retrieved evidence is insuffi- cient. Although our experiments are conducted on de-identified public benchmarks, any real world deployment would require strict privacy protection, external validation, calibration, subgroup fairness evaluation, and continuous clinician oversight. We therefore position D 2 MDT as a decision-support research framework rather than an autonomous di- agnostic or treatment system. Ethical Considerations This work uses MIMIC-I and MIMIC-IV, which are publicly available de-identified clinical datasets distributed through PhysioNet under controlled ac- cess. We use these datasets only for research on ICU outcome prediction and follow their data use requirements. We do not attempt to re-identify patients, redistribute the data, or use the data for clinical deployment or patient-level intervention. The proposed framework is intended as a research prototype for clinical decision support rather than an autonomous diagnostic or treatment system. References Erin P. Balogh. 2015. Improving diagnosis in health care. Savyasachi V. Shah Bds. 2024. Accuracy, consistency, and hallucination of large language models when analyzing unstructured clinical notes in electronic medical records. JAMA Network Open, 7(8):2. Ofir Ben Shoham and Nadav Rappoport. 2024. Cpllm: Clinical prediction with large language models. PLOS Digital Health, 3(12):e0000680. Alexandra Campione Russo, Jean-Luc Tilly, Leah Kauf- man, Melissa Danforth, Mark L. Graber, J. Matthew Austin, and Hardeep Singh. 2025. Hospital commit- ments to address diagnostic errors: An assessment of 95 us hospitals. Journal of Hospital Medicine, 20(2). Rochana Chaturvedi, Yue Zhou, Andrew Boyd, Brian T. Layden, Mudassir Rashid, Lu Cheng, Ali Cinar, and Barbara Di Eugenio. 2025. Early risk prediction with temporally and contextually grounded clinical language processing. Kai Chen, Xinfeng Li, Tianpei Yang, Hewei Wang, Wei Dong, and Yang Gao. 2025a. Mdteamgpt: A self- evolving llm-based multi-agent framework for multi- disciplinary team medical consultation. Xi Chen, Huahui Yi, Mingke You, WeiZhi Liu, Li Wang, Hairui Li, Xue Zhang, Yingman Guo, Lei Fan, Gang Chen, et al. 2025b. Enhancing diagnostic capabil- ity with multi-agents conversational large language models. NPJ digital medicine, 8(1):159. Edward Choi, Mohammad Taha Bahadori, Andy Schuetz, Walter F. Stewart, and Jimeng Sun. 2016a. Retain: Interpretable predictive model in healthcare using reverse time attention mechanism. Curran As- sociates Inc. Edward Choi, Mohammad Taha Bahadori, Jimeng Sun, Joshua Kulas, Andy Schuetz, and Walter Stewart. 2016b. Retain: An interpretable predictive model for healthcare using reverse time attention mechanism. Advances in neural information processing systems, 29. Xinsong Du, Zhengyang Zhou, Yifei Wang, Ya-Wen Chuang, Yiming Li, Richard Yang, Wenyu Zhang, Xinyi Wang, Xinyu Chen, Hao Guan, et al. 2026. Testing and evaluation of generative large language models in electronic health record applications: a systematic review. Journal of the American Medical Informatics Association, page ocaf233. Mark L Graber, Diana Rusz, Melissa L Jones, Diana Farm-Franks, Barbara Jones, Jeannine Cyr Gluck, Dana B Thomas, Kelly T Gleason, Kathy Welte, Jen- nifer Abfalter, et al. 2017. The new diagnostic team. Diagnosis, 4(4):225–238. Paul Hager, Friederike Jungmann, Robbie Holland, Ku- nal Bhagat, Inga Hubrecht, Manuel Knauer, Jakob Vielhauer, Marcus Makowski, Rickmer Braren, and Georgios Kaissis. 2024. Evaluation and mitigation of the limitations of large language models in clinical decision-making. Nature Medicine, 30(9):26. Hrayr Harutyunyan, Hrant Khachatrian, David C Kale, Greg Ver Steeg, and Aram Galstyan. 2019. Multitask learning and benchmarking with clinical time series data. Scientific data, 6(1):96. Stefan Hegselmann, Georg von Arnim, Tillmann Rheude, Noel Kronenberg, David Sontag, Gerhard Hindricks, Roland Eils, and Benjamin Wild. 2025. Large language models are powerful electronic health record encoders. arXiv preprint arXiv:2502.17403. Yixing Jiang, Kameron C. Black, Danny Park, James Zou, Andrew Y. Ng, and Jonathan H. Chen. 2025. Medagentbench: A realistic virtual ehr environment to benchmark medical llm agents. Alistair EW Johnson, Lucas Bulgarelli, Lu Shen, Alvin Gayles, Ayad Shammout, Steven Horng, Tom J Pol- lard, Sicheng Hao, Benjamin Moody, Brian Gow, et al. 2023. Mimic-iv, a freely accessible electronic health record dataset. Scientific data, 10(1):1. Alistair EW Johnson, Tom J Pollard, Lu Shen, Li-wei H Lehman, Mengling Feng, Mohammad Ghassemi, Benjamin Moody, Peter Szolovits, Leo Anthony Celi, and Roger G Mark. 2016. Mimic-i, a freely accessi- ble critical care database. Scientific data, 3(1):1–9. Misuk Kim and Kyu-Baek Hwang. 2022. An empirical evaluation of sampling methods for the classification of imbalanced data. PloS one, 17(7):e0271260. Yubin Kim, Chanwoo Park, Hyewon Jeong, Yik Siu Chan, Xuhai Xu, Daniel Mcduff, Hyeonhoon Lee, Marzyeh Ghassemi, Cynthia Breazeal, and Hae Won Park. 2024. Mdagents: An adaptive collaboration of llms for medical decision-making. Ching Yi Lee, Hung Yi Lai, Ching Hsin Lee, Mi Mi Chen, and Sze Yuen Yau. 2024. Collaborative clini- cal reasoning: a scoping review. PeerJ. Weibin Liao, Yinghao Zhu, Zhongji Zhang, Yuhang Wang, Zixiang Wang, Xu Chu, Yasha Wang, and Liantao Ma. 2024. Learnable prompt as pseudo- imputation: Rethinking the necessity of traditional ehr data imputation in downstream clinical predic- tion. Qicai Liu, Zhichao Hu, Tao Huang, Yupeng Niu, Xinche Zhang, Shanwu Ma, Chutong Lin, Goh Kim Huat, Hyeokkoo Eric Kwon, and Feng Gao. 2026. Evomdt: a self-evolving multi-agent system for structured clin- ical decision-making in multi-cancer. npj Digital Medicine, 9(1). Liantao Ma, Junyi Gao, Yasha Wang, Chaohe Zhang, and Xinyu Ma. 2020. Adacare: Explainable clin- ical health status representation learning via scale- adaptive feature extraction and recalibration. Pro- ceedings of the AAAI Conference on Artificial Intelli- gence, 34(1):825–832. Xinyu Ma, Yasha Wang, Xu Chu, Liantao Ma, Wen Tang, Junfeng Zhao, Ye Yuan, and Guoren Wang. 2022. Patient health representation learning via cor- relational sparse prior of medical features. IEEE Transactions on Knowledge and Data Engineering, 35(11):11769–11783. Matthew B McDermott, Haoran Zhang, Lasse H Hansen, Giovanni Angelotti, and Jack Gallifant. 2024. A closer look at auroc and auprc under class imbal- ance. Advances in Neural Information Processing Systems, 37:44102–44163. Harsha Nori, Mayank Daswani, Christopher Kelly, Scott Lundberg, Marco Tulio Ribeiro, Marc Wilson, Xi- aoxuan Liu, Viknesh Sounderajah, Jonathan Carlson, and Matthew P Lungren. 2025. Sequential diagnosis with language models. Robert S Porter and Justin L Kaplan. 2011. The Merck manual of diagnosis and therapy. Merck Sharp & Dohme Corp. Weijieying Ren, Jingxi Zhu, Zehao Liu, Tianxiang Zhao, and Vasant Honavar. 2025. A comprehensive survey of electronic health record modeling: From deep learning approaches to large language models. Andrew Sellergren, Sahar Kazemzadeh, Tiam Jaroensri, Atilla Kiraly, Madeleine Traverse, Timo Kohlberger, Shawn Xu, Fayaz Jamil, Cían Hughes, Charles Lau, et al. 2025. Medgemma technical report. arXiv preprint arXiv:2507.05201. Wenqi Shi, Ran Xu, Yuchen Zhuang, Yue Yu, Jieyu Zhang, Hang Wu, Yuanda Zhu, Joyce Ho, Carl Yang, and May D. Wang. 2024. Ehragent: Code empowers large language models for few-shot complex tabular reasoning on electronic health records. Mukund Sundararajan and Amir Najmi. 2020. The many shapley values for model explanation. In In- ternational conference on machine learning, pages 9269–9278. PMLR. Xiangru Tang, Anni Zou, Zhuosheng Zhang, Ziming Li, Yilun Zhao, Xingyao Zhang, Arman Cohan, and Mark Gerstein. 2024. Medagents: Large language models as collaborators for zero-shot medical reason- ing. Findings of the Association for Computational Linguistics ACL 2024, pages 599–621. Janneke E. W. Walraven, Olga L. Van, der Hel, J. J. M. Van, der Hoeven, Valery E. P. P. Lemmens, Rob H. A. Verhoeven, and Ingrid M. E. Desar. 2022. Factors influencing the quality and functioning of oncolog- ical multidisciplinary team meetings: results of a systematic review. BMC Health Services Research, 22(1):1–27. Haochun Wang, Sendong Zhao, Zewen Qiang, Nuwa Xi, Bing Qin, and Ting Liu. 2024. Beyond direct diag- nosis: Llm-based multi-specialist agent consultation for automatic diagnosis. Zixiang Wang, Yinghao Zhu, Huiya Zhao, Xiaochen Zheng, Dehao Sui, Tianlong Wang, Wen Tang, Yasha Wang, Ewen M. Harrison, Chengwei Pan, Junyi Gao, and Liantao Ma. 2025. Colacare: Enhancing elec- tronic health record modeling through large language model-driven multi-agent collaboration. In Proceed- ings of the ACM Web Conference 2025, pages 2250– 2261. Association for Computing Machinery. Ran Xu, Wenqi Shi, Yue Yu, Yuchen Zhuang, Bowen Jin, May Dongmei Wang, Joyce Ho, and Carl Yang. 2024. Ram-ehr: Retrieval augmentation meets clini- cal predictions on electronic health records. In Pro- ceedings of the 62nd Annual Meeting of the Associa- tion for Computational Linguistics (Volume 2: Short Papers), pages 754–765. Sihang Zeng, Lucas Jing Liu, Jun Wen, Meliha Yetis- gen, Ruth Etzioni, and Gang Luo. 2025. Trajsurv: Learning continuous latent trajectories from elec- tronic health records for trustworthy survival pre- diction. Rongjia Zhou, Chengzhuo Li, Carl Yang, and Jiaying Lu. 2025a. Clinnoteagents: An llm multi-agent sys- tem for predicting and interpreting heart failure 30- day readmission from clinical notes. Yucheng Zhou, Lingran Song, and Jianbing Shen. 2025b. Mam: Modular multi-agent framework for multi-modal medical diagnosis via role-specialized collaboration. In Findings of the Association for Computational Linguistics: ACL 2025, pages 25319– 25333. Yinghao Zhu, Ziyi He, Haoran Hu, Xiaochen Zheng, Xichen Zhang, Zixiang Wang, Junyi Gao, Liantao Ma, and Lequan Yu. 2025. Medagentboard: Bench- marking multi-agent collaboration with conventional methods for diverse medical tasks. Yinghao Zhu, Changyu Ren, Zixiang Wang, Xiaochen Zheng, Shiyun Xie, Junlan Feng, Xi Zhu, Zhoujun Li, Liantao Ma, and Chengwei Pan. 2024. Emerge: En- hancing multimodal electronic health records predic- tive modeling with retrieval-augmented generation. In Proceedings of the 33rd ACM International Con- ference on Information and Knowledge Management, pages 3549–3559. A Additional Method Details A.1 Semantic Evidence Construction For each expert, D 2 MDT computes feature attribu- tion scores on the latest-visit structured EHR input and ranks variables by their absolute contribution to the expert risk estimate. In the current implementa- tion, these attributions are produced by SHAP and saved as feature-level importance weights. D 2 MDT keeps the top important variables for each expert and verbalizes each retained variable with two ob- servable signals from the raw EHR sequence: its latest value and a short description of the local tem- poral trend near the latest observation time (e.g., persistent elevation, recent drop, abnormal fluctua- tion). This step converts model-specific numerical evidence into a compact textual card that can be in- terpreted by doctor agents without exposing latent vectors directly. Table 4 summarizes the retained fields. Table 4: Fields retained in D 2 MDT’s structured seman- tic and residual summaries. ComponentRetained fields Evidence cardlatest important variables, recent tempo- ral trends, and expert-specific risk cues derived from the EHR encoder and re- trieved knowledge Compact doctor card risk_vote,top_mechanisms, support_features, counter_evidence,confidence, need_escalation Routingsum- mary cross-expert abnormal findings, merged salient variables, and whether doctor risk votes are aligned or conflicting Compact leader report final_risk ,consensus_level, key_mechanisms,support_summary, minority_view, discussion_points A.2 Depatment routing Department assignment is implemented as a lightweight rule-based scorer. For each patient, D 2 MDT extracts three routing sources. The first is the abnormal numerical set fromraw_x. Con- cretely, for each continuous variable listed in Ta- ble 5, D 2 MDT compares the latest observed value against that patient’s own recent trajectory by com- puting a deviation score relative to the within- patient median and standard deviation; the variable is marked abnormal when this normalized deviation exceeds a preset threshold. The second source is the important-feature set. D 2 MDT reads the stored feature-importance file, ranks variables by the abso- lute SHAP-based importance weight, and keeps the top-kvariables for routing. The third source is the context-risk signal. This signal is not produced by a text encoder; instead, the code applies rule-based cue extraction to the patient context text and parses coarse indicators such as predicted mortality risk, whether the patient is in ICU, and whether the case falls into the top 30% risk range. For a department candidated, the routing score is the weighted sum of three parts: abnormal-numeric score, important-feature score, and context-risk score. The first two are accumulated through a feature-to-department match functionρ d (f). In the current implementation,ρ d (f)is computed by exact feature-to-department mapping when the vari- able name is one of the predefined D 2 MDT vari- ables in Table 5, and otherwise by keyword-overlap matching between the normalized variable name and the department keyword lexicon. The third term is a department-specific bonusc d triggered by the parsed risk cues. In the current implemen- tation, the three weights are 1.4, 0.9, and 0.7, re- spectively. The system computes scores for all can- didate departments, but the dynamic assignment step only ranks the non-fixed departments because the first doctor is always anchored to the critical- care/hemodynamic role. A.3 Prompt Template D 2 MDT uses a multi-stage prompt pipeline adapted to department-aware consultation and residual deliberation. We report example prompt templates below (placeholders in<>). The stages align with the initial consensus and later-round residual updates described in Section 2. A.4 Inference and Training The inference pipeline of D 2 MDT consists of four stages. First, three doctor agents independently an- alyze the same patient using different EHR expert backbones and their assigned departments. Second, the leader agent summarizes these initial opinions into a synthesized report and produces an initial risk judgment. Third, the doctor agents enter iter- ative consultation rounds, where they reconsider their opinions according to the leader report and the current disagreement points. Finally, the leader agent revises the final logits based on the completed consultation outputs. For prediction training, the collaboration out- puts are converted into fusion features and used to train a lightweight fusion model. The fusion model takes the EHR embeddings and text em- beddings produced by the collaboration stage as VariableImpute ValueModeled as Capillary refill rate-1categorical Glascow coma scale eye opening-1categorical Glascow coma scale motor response-1categorical Glascow coma scale total-1categorical Glascow coma scale verbal response-1categorical Diastolic blood pressurez-scorecontinuous Fraction inspired oxygenz-scorecontinuous Glucosez-scorecontinuous Heart Ratez-scorecontinuous Heightz-scorecontinuous Mean blood pressurez-scorecontinuous Oxygen saturationz-scorecontinuous Respiratory ratez-scorecontinuous Systolic blood pressurez-scorecontinuous Temperaturez-scorecontinuous Weightz-scorecontinuous pHz-scorecontinuous Table 5: The 17 selected clinical variables. The second column lists the imputation values used in preprocessing, and the third column describes how variables are modeled (categorical or continuous). input, and learns the final prediction under the same train/validation/test split protocol. B Additional Experimental Details B.1 Datasets and Splits We evaluate D 2 MDT on two widely used ICU out- come prediction benchmarks, MIMIC-I Out- come (Johnson et al., 2016) and MIMIC-IV Out- come (Johnson et al., 2023). Both datasets are derived from real-world intensive care unit elec- tronic health records (EHRs), and the task is in- hospital outcome prediction based on structured multivariate clinical time series. The input records contain longitudinal patient measurements together with basic clinical attributes, enabling the model to reason over temporal patient states under realistic clinical settings. The statistics of the dataset splits are summarized in Table 6. To enhance clinical reasoning, we additionally introduce the Merck Manual of Diagnosis and Therapy (MSD) (Porter and Kaplan, 2011) as an external medical guideline corpus. During infer- ence, D 2 MDT retrieves relevant MSD knowledge as supportive evidence for doctor agents, so that the multi-agent discussion is grounded not only in patient-specific EHR signals but also in explicit medical knowledge. This external corpus is used Table 6: Statistics of the experimented datasets after pre- processing. The # Samples column reports the number of patient records and their percentage in each split. DatasetSplit# Samples MIMIC-I Train19,026 (90.00%) Val1,057 (5.00%) Test1,057 (5.00%) MIMIC-IV Train13,302 (90.00%) Val739 (5.00%) Test739 (5.00%) only at inference time and does not alter the bench- mark definition or the training labels. B.2 Baseline categories We organize the compared methods into three groups according to their modeling assumptions and clinical decision-making mechanisms. Deep learning models. This group contains strong structured EHR prediction methods that directly model multivariate clinical time series without ex- plicit LLM-based collaboration. •AdaCare (Ma et al., 2020) is an interpretable clinical time-series model based on scale- adaptive feature extraction and recalibration. It Initial doctor prompt. System. An experienced doctor analyzes multivariate EHR, model-predicted mortality risk, feature importance, and retrieved medical knowledge. The patient record is prefixed with an assigned department hint. Inputs. Retrieved knowledgecontext; patient record and model analysis hcontext. Output JSON. Fields logit, analysis, and evidences. Derived compact doctor card. D 2 MDT builds a structured card with risk_vote, top_mechanisms, support_features, counter_evidence, confidence, and need_escalation. Initial leader synthesis. System. The leader reads all doctor opinions, checks whether they are clinically reasonable, and writes a synthesized patient-level report. Inputs. Patient basic information patient_info; aggregated doctor statements generated by generate_doctors_prompt(..., is_initial=True). Output JSON. Fields answer, report, and evidences. Derived compact leader report. The report is compressed into final_risk, consensus_level, key_mechanisms, support_summary, minority_view, and discussion_points. Later-round doctor revision. System. Each doctor revisits the leader conclusion and provides an updated view. Default inputs. Retrieved knowledge context; the doctor’s previous analysis analysis; leader opinion opinion; leader report report. Delta-discussion inputs. For roundsr≥2, D 2 MDT substitutes these slots with compact residual states: doctor compact text, disagreement points, and compact leader report. Output JSON. Fields answer (agree/disagree), confidence (1–3), reason, and evidences. Later-round leader update and stopping decision. Implemented summary prompt. Read the previous synthesized report plus all revised doctor statements, decides whether another round is needed, and updates the patient-level report. Inputs. Previous leader report latest_info; revised doctor statements fromgenerate_doctors_prompt(..., is_initial=False). Output JSON. Fields action, answer, report, and reasons. Auxiliary stopping prompt. A lighter stop/continue prompt output contains only action and reason. Figure 6: Implementation-faithful prompt and compact-state templates used in D 2 MDT. The first row shows the actual first-round doctor and leader prompts; the second row shows the later-round revision and stopping/update stages, including the compact semantic states injected by residual deliberation. captures patient status dynamics while highlight- ing clinically important variables, and is there- fore a representative deep learning baseline for EHR outcome prediction. •RETAIN (Choi et al., 2016a) is a reverse- time attention model for healthcare prediction. It is widely adopted as a strong interpretable EHR baseline because it assigns visit-level and variable-level attention weights while preserving competitive predictive performance. •PAI (Liao et al., 2024) is a recent EHR prediction framework that replaces explicit missing-value imputation with learnable prompts as pseudo- imputation. Since ICU EHR data contain substan- tial missingness, PAI serves as a strong recent baseline for evaluating whether D 2 MDT remains effective under modern missing-data-aware mod- eling. LLM-based models. This group introduces exter- nal medical knowledge or large language model reasoning into the prediction process, but does not explicitly implement a department-aware MDT col- laboration mechanism as D 2 MDT does. •EMERGE (Zhu et al., 2024) is a retrieval- augmented multimodal EHR framework that extracts medical entities from time-series data and clinical notes, aligns them with external medical knowledge, and generates task-relevant patient summaries for downstream prediction. We include EMERGE because it represents a strong RAG-style LLM-enhanced EHR predic- tion pipeline. •MedGemma Zero-Shot and MedGemma Few- Shot (Sellergren et al., 2025) use the MedGemma medical foundation model directly for clinical reasoning under zero-shot and few-shot prompt- ing settings, respectively. These baselines are used to test whether general-purpose medical foundation models can be directly transferred to structured ICU outcome prediction without additional task-specific collaboration design. MDT-based models. This group is the closest com- Algorithm 1 Training and Inference of D 2 MDT Require: Processed EHR datasetD, specialist encoder set M =AdaCare, MCGRU, RETAIN, LLM-based collaboration moduleC Ensure: Final prediction score ˆy for each patient Stage I: Train specialist EHR encoders 1: for each encoder m∈M do 2:Train m on the training split ofD 3:Select the best checkpoint according to validation AUPRC 4:Use the best checkpoint to extract patient embeddings on validation and test splits 5: end for Stage I: Construct fusion data 6: for each patient p do 7:Collect specialist EHR embeddings e (m) p m∈M 8:Generate the final collaborative report r p usingC 9:Encode r p into text embedding t p 10:Build fusion sample (e (m) p , t p ,y p ) 11: end for 12: Split fusion data into train/validation/test sets 13: Train the fusion predictor with binary cross-entropy loss 14: Select the best fusion checkpoint according to validation AUPRC Inference 15: for each test patient p do 16: Obtain specialist embeddingse (m) p m∈M 17:Generate collaborative report r p and encode it as t p 18: Fusee (m) p andt p to obtain predictionˆy p 19: end for parison family to D 2 MDT, since all methods ex- plicitly simulate multi-doctor consultation or multi- agent medical decision-making. • MDAgents (Kim et al., 2024) is an adaptive med- ical multi-agent framework that dynamically as- signs collaboration structures among LLM agents for medical decision-making. It serves as a rep- resentative general-purpose medical multi-agent baseline. • MDTeamGPT (Chen et al., 2025a) is a self- evolving LLM-based MDT consultation frame- work with consensus aggregation, residual dis- cussion structure, and experience accumulation. We compare against MDTeamGPT because it em- phasizes multi-round consultation efficiency and knowledge reuse, which are closely related to the design goals of D 2 MDT. •ColaCare (Wang et al., 2025) is the most direct baseline to D 2 MDT. It integrates EHR expert models, LLM doctor agents, a meta-agent, and MSD-based retrieval within an MDT-style collab- orative consultation framework. Compared with ColaCare, D 2 MDT further introduces patient- specific department assignment and compact disagreement-focused discussion, so this com- parison directly evaluates the contribution of our design. Whenever a baseline is reproduced locally, we use the same dataset split, evaluation pipeline, and prediction post-processing as D 2 MDT. This con- trolled setup keeps the comparison focused on modeling and collaboration differences rather than implementation-side evaluation discrepancies. B.3 Metrics details AUROC (McDermott et al., 2024): This metric is our primary consideration in binary classifica- tion tasks due to its widespread use in clinical set- tings and its effectiveness in handling imbalanced datasets. AUPRC (Kim and Hwang, 2022): The AUPRC is particularly useful for evaluating perfor- mance in datasets with a significant imbalance be- tween classes. min(+P, Se) (Ma et al., 2022): This composite metric represents the minimum value between precision (+P) and sensitivity (Se), provid- ing a balanced measure of model performance. All these three metrics are the higher the better. AR1PT andAR2PT as the average prompt to- kens consumed in round 1 and round 2, respectively. AR2PT being smaller than AR1PT indicates that residual discussion is effective. B.4 Implementation stack. D 2 MDT uses three doctor agents and one leader agent. The three doctor agents are paired with AdaCare, MCGRU, and RETAIN, respectively, while the leader agent summarizes specialist opin- ions and produces the final consensus report. In the full D 2 MDT setting, the consultation pipeline enables department-aware prompting, department- specific retrieval, structured doctor cards, semantic doctor cards, compact leader reports, and residual disagreement discussion. All consultation agents are driven by Qwen3-8B. For retrieval-augmented consultation, we use MedCPT together with the MSD corpus. 2345 max_round 84.4 84.8 85.2 85.6 AUROC (%) (a) MIMIC-I AUROC 2345 max_round 85 86 87 88 AUROC (%) (b) MIMIC-IV AUROC Figure 7: Sensitivity analysis of AUROC under different values of max_round. For multimodal prediction, the final leader re- port is encoded by GatorTron-base. Each report is tokenized with a maximum length of 512 and repre- sented by the last-layer[CLS]embedding, yielding a 1024-dimensional text feature. This text feature is concatenated with the three 128-dimensional EHR embeddings from AdaCare, MCGRU, and RETAIN. The fusion module is a lightweight MLP that first projects the concatenated multimodal vec- tor to a 128-dimensional hidden representation, ap- plies a GELU activation, and then outputs the final risk probability through a linear layer followed by a sigmoid. We train the fusion model with AdamW using a learning rate of1× 10 −3 and binary cross-entropy loss. The maximum number of training epochs is 50, the early-stopping patience is 10, and model selection is based on validation AUPRC. The batch size is 128. We fix the random seed to 42 for re- producibility. Unless otherwise specified, the max- imum number of consultation rounds is set to 3 for both MIMIC-I and MIMIC-IV. B.5 Hardware and software environment. The current local environment provides multiple NVIDIA GeForce RTX 4090 24GB GPUs (driver version 590.48.01). The repository requirements specify torch 2.3.1, lightning 2.3.3, transformers 4.42.4, numpy 1.26.4, pandas 2.2.2, scikit_learn 1.5.1, shap 0.46.0, faiss_cpu 1.8.0.post1, and faiss_gpu 1.7.2. In practice, the exact Python run- time and serving backend may vary across ma- chines, but all reported experiments follow the same repository-level configuration, retrieval stack, and evaluation scripts. B.6 Sensitivity Analysis Figure 7, Figure 8, Figure 9 and Figure 10 report the remaining sensitivity results that are not shown in the main text. These figures complement the main-text AUPRC analysis by showing how pre- dictive robustness and communication cost change as the maximum consultation round varies. 2345 max_round 52 53 54 55 min(+P, Se) (%) (a) MIMIC-I min(+P, Se) 2345 max_round 60.0 61.5 63.0 min(+P, Se) (%) (b) MIMIC-IV min(+P, Se) Figure 8: Sensitivity analysis of min(+P, Se) under dif- ferent values of max_round. 2345 max_round 8140 8155 8170 8185 8200 AR1PT (a) MIMIC-I AR1PT 2345 max_round 9075 9100 9125 9150 9175 AR1PT (b) MIMIC-IV AR1PT Figure 9: Sensitivity analysis of communication cost AR1PT under different values of max_round. 2345 max_round 6400 6800 7200 7600 AR2PT (a) MIMIC-I AR2PT 2345 max_round 6600 6900 7200 7500 AR2PT (b) MIMIC-IV AR2PT Figure 10: Sensitivity analysis of predictive metrics under different values of min(+P, Se). B.7 Case Study Case study of a patient, illustrating how the dis- cussion evolves from the initial consensus to the final report. Based on the initial consensus, the leader-level synthesis initially assessed the patient as low risk, emphasizing hemodynamic instability, perfusion compromise, and oxygenation compro- mise. In the subsequent round 1 review, however, Doctor 1 disagreed with this low-risk conclusion, arguing that the AdaCare score of 0.19, the pa- tient’s top-30% risk ranking among ICU patients, and the possibility of physiological instability as- sociated with long-term renal replacement therapy, metabolic derangements, and acid-base disturbance together suggested a higher mortality risk. Doc- tors 2 and 3, in contrast, agreed with the low-risk synthesis, considering the observed abnormalities insufficient on their own to justify a high-risk inter- pretation. After aggregating these responses, the leader produced the round1 summary, which served as the effective initial consensus for the next stage. At this point, the leader revised the conclusion to high risk, identified hemodynamic instability, oxy- genation compromise, and metabolic stress as the key mechanisms, and explicitly marked the remain- ing minority disagreement as the focus for further discussion. Because this unresolved disagreement persisted, the case proceeded to round 2. In round 2, the system forwarded only compact structured summaries and the residual disagreement point, rather than replaying the full first-round discus- sion. Under this focused setting, all three doc- tors converged on a high-risk judgment. Doctor 1 emphasized chronic renal tubular acidosis and life-threatening electrolyte disturbances, Doctor 2 highlighted oxygenation compromise and respira- tory failure, and Doctor 3 stressed the metabolic complications of chronic RTA and long-term dial- ysis. The final leader report therefore concluded that the patient had a high risk of mortality with strong consensus and no remaining minority view, showing that the second-round residual discussion helped resolve the initial disagreement and refine the final clinical assessment.

Instant research discovery