The paper introduces AegisDx, a safety-oriented framework for AI-assisted differential diagnosis. It coordinates role-specialized LLM components with contracts, structured intermediate outputs, evidence retrieval, and verification gates. Using GPT-oss-120B, AegisDx improved Top-3 diagnostic accuracy over a standalone model on JAMA, NEJM, and Annals of Emergency Medicine cases, while capturing more physician-consensus “must-not-miss” conditions. In a blinded evaluation of 43 real-world emergency-department notes, physicians rated its composite safety score at 4.55 versus 4.31 for GPT-5, with an adjusted p-value of 2.1×10^-4.
No heat snapshots are available in the last 24 hours.