EHR-R1: A Reasoning-Enhanced Foundational Language Model for Electronic Health Record Analysis

This repository contains the EHR-Ins-Reasoning dataset, as presented in the paper EHR-R1: A Reasoning-Enhanced Foundational Language Model for Electronic Health Record Analysis.

EHR-Ins is a large-scale, comprehensive instruction dataset developed to enhance the reasoning and analysis capabilities of Large Language Models (LLMs) for Electronic Health Records (EHR).

• Composition and Scale: It is a instruction corpus that comprises two major types of data:

  • High-Quality Reasoning Cases: 300K reasoning cases.
  • Non-Reasoning Cases: Approximately 3.5 to 4 million non-reasoning cases.

• Task Coverage: The dataset spans a wide variety of 42 distinct EHR tasks, categorized into two types: decision-making (e.g., diagnosis and treatment recommendations) and risk-prediction (e.g., mortality and readmission).

• Core Innovation (Thinking-Graph): The dataset's core innovation is a "thinking-graph-driven framework" used to synthesize the high-quality reasoning data at scale. This pipeline works by:

  • Identifying key related medical entities from EHRs.
  • Linking these entities using external knowledge, such as the UMLS knowledge base.
  • Prompting a model (like GPT-4o) to produce structured, step-by-step clinical reasoning based on the generated graph.

• Purpose: EHR-Ins provides explicit medical reasoning supervision, which enables models like the EHR-R1 series to systematically acquire diverse, context-rich reasoning capabilities necessary for accurate and robust EHR analysis.

• Paper: https://huggingface.co/papers/2510.25628

• GitHub Repository: https://github.com/MAGIC-AI4Med/EHR-R1

Structure

Each item in the jsonl file contains the key as below:

• idx: Unique ID for each sample
• instruction: Task instruction; the instruction is the same if the task is the same
• input: EHR input after text serialization
• output: Output used for training (this item is not useful for the test set)
• candidates: Candidate options provided for the untrained model
• task_info: Task-related information is included in this item, including:
  • target_key: The column name from the EHR used to retrieve the prediction label; this item is None for the risk prediction task
  • events: Event types related to the prediction label
  • metric: The metric used to calculate the score for this task
  • target: The raw label in string format
  • label: The label used to calculate the score

To prevent the leakage of native data information within the MIMIC-IV dataset, we removed information such as subject_id, harm_id, and other details that might link to the original MIMIC-IV data. The complete dataset can be found in MIMIC-IV-Ext-EHR-Analysis on PhysioNet (not yet released).

📖 Citation

If you find our work helpful or inspiring, please feel free to cite it:

@article{liao2025ehrr1,
  title={{EHR-R1: A Reasoning-Enhanced Foundational Language Model for Electronic Health Record Analysis}},
  author={Liao, Yusheng and Wu, Chaoyi and Liu, Junwei and Jiang, Shuyang and Qiu, Pengcheng and Wang, Haowen and Yue, Yun and Zhen, Shuai and Wang, Jian and Fan, Qianrui and Gu, Jinjie and Zhang, Ya and Wang, Yanfeng and Wang, Yu and Xie, Weidi},
  journal={arXiv preprint arXiv:2510.25628},
  year={2025}
}