🧠 MIMIC-SR-ICD11: A Dataset for Narrative-Based Diagnosis

Authors: Yuexin Wu (University of Memphis), Shiqi Wang (Guangzhou University of Chinese Medicine), Vasile Rus (University of Memphis) Conference: Findings of ML4H 2025 Paper: MIMIC-SR-ICD11: A Dataset for Narrative-Based Diagnosis

---

🌍 Overview

MIMIC-SR-ICD11 is a large-scale English diagnostic dataset that converts de-identified EHR discharge notes from MIMIC-IV into patient-authored self-reports and standardizes diagnoses with the WHO ICD-11 ontology. It bridges clinical documentation and real-world patient narratives, enabling diagnostic reasoning on natural-language symptom descriptions rather than categorical checklists.

• Inputs mirror first-contact symptom narratives (patient self-reports).
• Labels are natively aligned to ICD-11, reducing ambiguity and post-hoc mapping.
• Designed for full-text diagnostic inference and realistic downstream deployment (triage tools, assistants, conversational agents).

---

🧱 Dataset Construction

This repository displays the construction as Figure 2 from the paper.

Figure 2: Left branch maps primary diagnoses from MIMIC-IV (ICD-9 → ICD-10 via CMS GEMs; ICD-10 → ICD-11 via WHO tables) with one-to-one filtering and manual curation. Right branch rewrites MIMIC-IV-Note into first-person self-reports using an instruction-tuned prompt that excludes clinician-only content.

---

🧮 LL-Rank: PMI-Style Re-Ranking

We introduce LL-Rank, a likelihood-based re-ranking framework that discounts head-class (frequent label) bias by subtracting a report-free prior from the conditional likelihood.

$$ S(x, c) = -L_{\text{cond}}(x,c) + \alpha L_{\text{prior}}(c), \quad \alpha = 1 $$

• Conditional term $L_{\text{cond}}$: per-token NLL of label c given report x under a fixed prompt.
• Prior term $L_{\text{prior}}$: per-token NLL of c under the same prompt without the report (report-free prior).
• Effect: isolates semantic compatibility from label frequency, improving calibration and long-tail performance.

---

📈 Results Highlights

Across seven medical backbones, LL-Rank consistently outperforms a generation+mapping baseline (GenMap):

• Hit@3/5/10: average gains of **≈ +80% / +86% / +93%
• Macro-F1@3/5/10: average gains of **≈ +138% / +147% / +157%

These gains are especially strong for Macro-F1, indicating improved performance on underrepresented labels—not merely amplifying head classes. Performance peaks around (\alpha \approx 1).

---

🧪 Baselines (Summary)

---

# Clone
git clone https://github.com/woqingdoua/MIMIC-SR-ICD11.git
cd MIMIC-SR-ICD11

# (Optional) Setup virtual environment
# python -m venv .venv && source .venv/bin/activate

# Install project dependencies (example)
pip install -r requirements.txt

# Configure Hugging Face auth securely (choose ONE)
# 1) Local keyring-based login (recommended for laptops)
huggingface-cli login
# 2) Environment variable (recommended for servers/CI)
# export HF_TOKEN=***

---

📚 Citation

If you use this repository, please cite the paper:

@inproceedings{wu2025mimicsricd11,
  title     = {MIMIC-SR-ICD11: A Dataset for Narrative-Based Diagnosis},
  author    = {Wu, Yuexin and Wang, Shiqi and Rus, Vasile},
  booktitle = {Findings of Machine Learning for Health (ML4H)},
  year      = {2025},
  url       = {https://github.com/woqingdoua/MIMIC-SR-ICD11}
}

---

🩺 Contact

Yuexin Wu · ywu10@memphis.edu