RareArena: 50,000 Patient Summaries Across 4,000 Rare Diseases

RareArena contains diagnostic data for approximately 50,000 patients across more than 4,000 rare diseases, developed by THUMedInfo and last updated in early 2026. The dataset is derived from the PMC-Patients corpus of clinical case reports and utilizes GPT-4o for the extraction of ground-truth diagnoses.

Use Cases

• Evaluating the diagnostic accuracy of medical LLMs using the ground-truth diagnosis labels
• Training question-answering systems to identify rare conditions from patient summaries
• Benchmarking clinical information extraction pipelines on complex rare disease case reports

Strengths

• Covers over 4,000 distinct rare diseases
• Includes nearly 50,000 patient records
• Sourced from peer-reviewed PMC case reports

Limitations

• Data extraction and processing were performed by GPT-4o, which may introduce model-specific biases or extraction errors
• Geographic bias toward the United States as indicated by dataset metadata tags

Provenance

Source

PMC-Patients (sourced from PubMed Central case reports)

Collection Method

Extracted and processed from PMC-Patients using GPT-4o

Freshness

Last updated February 2026.

Geography

United States