EHRWorld: A Patient-Centric Medical World Model for Long-Horizon Clinical Trajectories

1 Introduction

• Although LLMs perform well on static medical tasks, they struggle to maintain consistent patient states during sequential interventions, leading to accumulated errors in multi-step simulations.The introduction also motivates trajectory modeling by the need to reason about outcomes under alternative treatment plans for personalized care.Source: §1 · p. 2
• EHRWorld-110K captures complete patient-care trajectories from admission to discharge, providing longitudinal data for learning patient-state evolution and intervention-conditioned transitions.The dataset comprises approximately 110,000 hospitalization episodes and 17.5 million clinical events.Source: §1 · p. 2
• EHRWorld models clinical trajectories as a continuous sequential process, learning intervention-driven physiological transitions and updating patient states in response to clinical actions.The model is presented as a family of patient-centric medical world models trained under a causal sequential paradigm.Source: §1 · p. 2
• EHRWorld significantly outperforms naive LLM-based baselines in long-horizon simulation, with reduced error accumulation, improved stability on clinically sensitive events, and favorable reasoning efficiency.The authors present these as results of their comprehensive evaluation.Source: §1 · p. 2

2.1 Naive LLM-based Simulation

• Prompt-based virtual-patient systems support multiturn clinical interaction, but fixed case descriptions leave patient states static and do not model intervention-driven disease progression over time.The related-work discussion names AgentClinic, AutoMedic, CP-Env, and MAQuE as representative systems.Source: §2.1 · p. 3

2.2 EHR World Models

• BEHRT and Med-BERT use longitudinal histories for risk estimation, but predict from fixed sequences rather than modeling state changes caused by clinical interventions.These models are characterized as passive observers in this discussion.Source: §2.2 · p. 3
• Prior synthetic trajectory methods typically do not support state transitions conditioned on therapeutic actions, whereas EHRWorld models patient-state evolution as a causal, dynamic process under sequential interventions.The cited contrast includes rule-based Synthea and data-driven PatientSim.Source: §2.2 · p. 3

3.1 Data Sources and Preprocessing

• Discharge summaries are parsed for demographics and hierarchical diagnoses, while time-stamped events are organized into hospital-stay sequences and classified as inquiries or interventions.Inquiries yield explicit values; interventions aim to modify patient state without directly generating observable outcomes.Source: §3.1 · p. 3

3 Data Construction Pipeline

• The pipeline constructs EHRWorld-110K from MIMIC-IV records by integrating static patient context with longitudinal clinical events at the hospitalization-episode level.The figure overview depicts the data-construction pipeline alongside the model’s sequential simulation loop.Source: §3 · p. 3, §3.1 · p. 3, Figure 2 · p. 4

3.2 Data Filtering and Partition

• After quality filtering, EHRWorld-110K contains 110,513 hospitalization episodes and approximately 17.5 million clinical events, with episode events ordered chronologically.Admissions with insufficient clinical activity or incomplete records are excluded.Source: §3.2 · p. 4
• The held-out test set contains 579 episodes, 84,010 inquiry events, 25,798 intervention events, and 1,043 unique primary and secondary diagnostic conditions.The remaining episodes are used for model training.Source: §3.2 · p. 4

4 The EHRWorld Model

• EHRWorld formulates patient simulation as a conditional world model that explicitly tracks patient states and updates them under sequential inquiries and interventions.Simulation steps correspond to physiological timestamps, and actions may occur concurrently.Source: §4 · p. 4

4.1 Patient State Representation

• The patient state includes a physiological timestamp, static demographics, an episode-level diagnostic profile, and the interaction history of prior clinical events.The diagnostic profile is extracted from discharge summaries and supplies episode-level clinical context.Source: §4.1 · pp. 4–5

4.2 Set-Based Conditional Generation

• At each step, the model conditions on a set of concurrent clinical orders and generates corresponding outcomes for those actions from a shared patient state.The set-based formulation is intended to capture clinical actions that influence patient states in parallel.Source: §4.2 · p. 5

4.3 Dual-Mode Prediction Mechanism

• EHRWorld uses dual-mode prediction: inquiries such as tests or examinations produce explicit observed values, while interventions are recorded without an immediate observable outcome.Intervention events update simulation context and can influence future clinical predictions.Source: §4.3 · p. 5

4.4 Deterministic State Transition

• After outcomes are generated, the interaction history is deterministically updated so subsequent predictions condition on prior inquiries and interventions.The passage describes this history update as supporting consistent long-horizon simulation of disease progression under clinical decision-making.Source: §4.4 · p. 5

5.1 Experimental Setup

• The experiment fine-tunes EHRWorld-4B, -8B, and -14B on EHRWorld-110K and compares them with closed-source, general-purpose, and medical LLM baselines.Evaluation covers continuous-variable S@k and SMAPE, clinical-status F1, and classification precision, recall, and F1; baselines use default inference configurations without task-specific fine-tuning.Source: §5.1 · pp. 5–6

5.2 Main Results: Full Trajectory Prediction

• EHRWorld achieves Avg Scores of 0.755–0.765 versus 0.618 for GPT-5.2 and 0.574 for Gemini-3.0-Pro-Preview in full-trajectory simulation.The full-trajectory comparison is reported in Table 1; the paper also reports lower Avg Scores of 0.401–0.506 for medical baselines MedGemma and Baichuan-M2.Source: Table 1 · p. 6, §5.2 · p. 7
• Scaling Qwen3-30B-A3B-Instruct to Qwen3-235B-A22B-Instruct raises S@25 only from 0.645 to 0.647, while EHRWorld-4B reaches 0.703 and exceeds Gemini-3.0-Pro-Preview at 0.681.The authors interpret this contrast as evidence that causal sequential training matters more than parameter scaling for patient-dynamics modeling.Source: §5.2 · p. 7

5.3 Next-Step vs. Full Trajectory Prediction

• EHRWorld-14B retains 92.6% overall performance from next-step to full-trajectory prediction, versus 86.2% for the best baseline, GPT-5.2.Next-step performance gaps are narrow, with Numer. S@25 ranging from 0.752 to 0.806, while the full-trajectory gap widens.Source: §5.3 · pp. 7–8, Table 2 · p. 7

5.4 High-Sensitivity Stability Analysis

• EHRWorld-14B’s S@25 falls from 71.6% to 69.7% (−1.9 points) and SMAPE rises by +0.030 (0.262 → 0.292) on high-sensitivity shifts.The subset comprises consecutive steps with relative change exceeding 50%; Figure 3 compares degradation on global and high-sensitivity data.Source: Figure 3 · p. 7, §5.4 · p. 8

5.5 Case Study

• In the case study, most EHRWorld-14B predictions are precise (relative error ≤10%), while dynamic Hematocrit predictions remain acceptable (10–20%).Figure 4 compares ground truth with predictions for a 69-year-old patient with left femoral artery stenosis.Source: §5.5 · p. 8, Figure 4 · p. 8

• Models incorporating medical knowledge struggle with state inconsistency and error drift under sequential interventions.Source: §6 · pp. 8–9
• EHRWorld and EHRWorld-110K support causally aligned clinical-trajectory modeling, and EHRWorld outperforms proprietary and open-source baselines in long-horizon prediction.Source: §6 · p. 9

• Autoregressive generation can propagate residual inaccuracies over extended rollouts despite reducing long-horizon error accumulation.Source: Limitations · p. 9
• The evaluation focuses on trajectory fidelity, stability, and consistency, not treatment optimization, policy learning, or real-world decision support.Source: Limitations · p. 9

Ethical Considerations

• Observational clinical data may reflect existing practice patterns and implicit biases that influence simulated outcomes if interpreted without caution.The authors state that EHRWorld-110K uses de-identified data processed under privacy and data-protection standards.Source: Ethical Considerations · p. 9
• EHRWorld is intended for research simulation and analysis, not autonomous or real-time clinical decision-making; broader use requires validation, transparency, and human oversight.Source: Ethical Considerations · p. 9

A Details on Data Construction

• EHRWorld-110K is assembled by extracting episode-level profiles, ordering event sequences, integrating them at the hospitalization level, and applying quality filtering.Source: A Details on Data Construction · p. 11

A.1 Static Information Extraction

• Static extraction from MIMIC-IV discharge summaries captures age, gender, allergy history, chief complaint, and admission-relevant diagnoses in a standardized schema.Source: §A.1 · p. 11

Static Information Extraction

• The extraction instructions distinguish primary from secondary diagnoses and exclude historical conditions unless actively treated during the stay.Source: §A.1 · p. 11

A.2 Data Filtering and Partitioning

• Filtering excludes episodes with metrics above the 90th percentile, event counts outside the 10th–90th percentile range, or no medication-administration records; patient-level splitting keeps each subject in one partition.Source: §A.2 · p. 11

A.3 Data Statistics and Distribution

• The dataset centers on patients aged 50–60, while most stays last 1–3 days and a long tail exceeds 100 days.The mean length of stay is approximately 5 days, versus a maximum above 100 days.Source: §A.3 · pp. 11–12
• Laboratory and medication events are most numerous, procedural events are sparse, and event intensity normalizes total events by length of stay.High-intensity episodes are often associated with longer hospitalizations.Source: §A.3 · p. 12

B EHRWorld Simulation Algorithm

• The simulation algorithm updates patient states over discrete steps in response to clinical actions, using dual-mode prediction for inquiry and intervention events.It is presented as an execution-flow description, not as an additional modeling assumption.Source: B EHRWorld Simulation Algorithm · p. 13

C Detailed Evaluation Metrics

• The evaluation framework separates numerical precision for continuous variables from semantic accuracy for discrete clinical events.The numerical metrics include Success Rate at threshold X and SMAPE, which is designed for symmetric penalties and numerical stability.Source: C Detailed Evaluation Metrics · p. 13, §C.1 · p. 13

C.1 Numerical Metrics

• Success Rate measures the share of predictions within a clinically acceptable relative-error threshold; higher values indicate greater accuracy.The reported thresholds are S@10, S@15, and S@25.Source: §C.1 · p. 13

C.2 High-Sensitivity Analysis

• High-sensitivity events are substantial ground-truth deviations from the preceding time step, with S@25 and SMAPE used to assess transitions above 50%.The analysis targets acute changes that global aggregate metrics may obscure.Source: §C.2 · p. 13

C.3 Label Event Metrics

• Discrete clinical outputs, including microbiology results and nursing assessments, are evaluated for semantic correctness using classification metrics based on precision and recall.Precision, recall, and F1 are used for categorical outputs.Source: §C.3 · p. 14

D.1 Inference Efficiency

• 1.29 seconds per case and 23,318 tokens/second give EHRWorld-14B the lowest latency and highest throughput among the compared models.It is faster than MedGemma-27B-IT at 2.60 seconds per case and exceeds GPT-OSS-120B at 14,571 tokens/second throughput.Source: §D.1 · p. 14

D.2 Hyperparameter Sensitivity

• Performance plateaus at 3 epochs, while further training yields minimal gains and increases computational cost.The reported optimal learning rate is 1 × 10−6; higher rates cause instability and lower rates slow convergence.Source: §D.2 · p. 14