MedCase-Bench

Project page · MedCase-150K · Agent code

MedCase-Bench contains 584 cases for medical case-report generation, with one clinical-atom JSON file per case and source bibliography. Clinical atoms are patient-specific facts extracted from published case reports, covering history, presentation, examinations, treatment and outcome.

Each atom file contains five lists of strings:

Field	Contents
history	Patient background and relevant medical history.
presentation	Symptoms and findings at presentation.
diagnostics	Examinations, laboratory results and diagnostic observations.
management	Treatments, procedures and clinical decisions.
outcome	Clinical course, response and follow-up.

Files

File	Contents
atoms/<PMCID>_atoms.json	584 clinical-atom files.
metadata.json	584 corresponding source records, including article identifiers, publication dates, licenses and source download links.

This package contains textual inputs and source metadata. Full articles and medical images can be obtained through the source links in metadata.json. Source publications retain the licenses recorded in their metadata.

Use

import json
from pathlib import Path

files = sorted(Path("atoms").glob("*_atoms.json"))
assert len(files) == 584
atoms = json.loads(files[0].read_text(encoding="utf-8"))
print(atoms["history"])

After installing MedCaseAgent and configuring a compatible model API:

medcase-agent validate atoms/<PMCID>_atoms.json
medcase-agent generate atoms/<PMCID>_atoms.json \
  --exclude-ids path/to/your-exclusions.json --output runs

For generation with images, prepare each case's source images following the preprocessing guide. For retrieval-based evaluation, prepare your own PMCID, PMID and DOI exclusion list for the benchmark sources and any related publications.