[论文解读] EHRWorld: A Patient-Centric Medical World Model for Long-Horizon Clinical Trajectories
EHRWorld 是一个因果序列化的医疗世界模型,基于大规模的 EHR 派生数据集(EHRWorld-110K)进行训练,能够在干预下实现对患者轨迹的稳定长期仿真,优于天真的大语言模型基线。
World models offer a principled framework for simulating future states under interventions, but realizing such models in complex, high-stakes domains like medicine remains challenging. Recent large language models (LLMs) have achieved strong performance on static medical reasoning tasks, raising the question of whether they can function as dynamic medical world models capable of simulating disease progression and treatment outcomes over time. In this work, we show that LLMs only incorporating medical knowledge struggle to maintain consistent patient states under sequential interventions, leading to error accumulation in long-horizon clinical simulation. To address this limitation, we introduce EHRWorld, a patient-centric medical world model trained under a causal sequential paradigm, together with EHRWorld-110K, a large-scale longitudinal clinical dataset derived from real-world electronic health records. Extensive evaluations demonstrate that EHRWorld significantly outperforms naive LLM-based baselines, achieving more stable long-horizon simulation, improved modeling of clinically sensitive events, and favorable reasoning efficiency, highlighting the necessity of training on causally grounded, temporally evolving clinical data for reliable and robust medical world modeling.
研究动机与目标
- Motivate the need for dynamic, causally grounded medical world models to simulate long-horizon patient trajectories under interventions.
- Create a large-scale, high-quality longitudinal dataset (EHRWorld-110K) from real-world EHRs to learn temporal dynamics and intervention-conditioned transitions.
- Develop EHRWorld, a family of patient-centric world models trained with a causal sequential paradigm to maintain consistent patient states over long horizons.
- Demonstrate that EHRWorld reduces error accumulation, improves stability during clinically sensitive events, and offers favorable reasoning efficiency compared to baseline LLMs.
提出的方法
- Formalize patient state as S_t = <τ_t, d, Y, H_t> where Y includes primary/secondary diagnoses with rationales.
- Model interaction as set-based conditional generation with actions A_t and outcomes V_t.
- Use a dual-mode prediction mechanism: inquiry actions generate observable values; intervention actions have empty immediate outcomes and update history.
- Train with a causal masking objective to learn intervention-conditioned state transitions and longitudinal dynamics.
- Construct EHRWorld-110K by extracting episode-level static context and temporally ordered event sequences from MIMIC-IV, followed by filtering and partitioning.
实验结果
研究问题
- RQ1Can LLMs with medical knowledge serve as dynamic medical world models for long-horizon clinical trajectories?
- RQ2Does training on causally grounded, temporally evolving clinical data (EHRWorld-110K) improve long-horizon state consistency and reduce error accumulation?
- RQ3How do EHRWorld models compare to naive LLM baselines in full trajectory prediction and stability under sequential interventions?
- RQ4What is the impact of model scale within the EHRWorld family on long-horizon clinical simulation performance?
- RQ5How robust is the model to high-sensitivity clinical transitions and abrupt physiological changes?
主要发现
- EHRWorld outperforms naive LLM baselines in long-horizon clinical trajectory simulation, with reduced error accumulation.
- EHRWorld achieves higher retention (stability) in full trajectory generation than GPT-5.2 and other baselines (e.g., 92.6% overall retention for EHRWorld-14B).
- Causal sequential training yields larger gains than mere parameter scaling for long-horizon robustness; EHRWorld-14B attains the best overall Avg score (0.765) among listed models.
- EHRWorld-14B provides strong next-step and full-trajectory performance, maintaining low relative error and high clinical interpretability across clinically sensitive events.
- High-sensitivity stability analysis shows EHRWorld-14B exhibits smaller performance degradation under abrupt state changes compared to general LLMs.
- Case study demonstrates EHRWorld maintains state consistency and accurate trajectory predictions across dynamic clinical scenarios.
阅读本论文
在 Nubint 阅读器中打开原文,边读边就任意段落提问。
在 Nubint 中阅读本论文本解读由 AI 根据论文原文生成,引用前请核对原文。