NarrativeShield: SDoH Persona-Varied Clinical Questions

NarrativeShield is a health equity benchmark for clinical NLP containing 1,000 USMLE-style clinical questions. Each question is instantiated across three sociolinguistic persona voices grounded in Social Determinants of Health literature, yielding 3,000 total clinical encounters. The dataset was created by Prabhjotschugh and was last updated on 2026-05-11.

Use Cases

• Benchmarking clinical NLP model fairness based on sociolinguistic persona variations
• Training models for equitable clinical question answering based on USMLE-style questions
• Studying the impact of Social Determinants of Health on clinical narratives

Strengths

• Contains 1,000 distinct USMLE-style clinical questions
• Provides 3,000 total clinical encounters through persona variations
• Explicitly grounded in Social Determinants of Health literature

Limitations

• Column-level documentation is absent; field semantics must be inferred after download
• Row count beyond the stated 3,000 encounters is unknown, which may limit suitability assessment

Provenance

Source

Prabhjotschugh via Hugging Face

Freshness

Last updated 2026-05-11 11:19:54