New AI Framework Enhances Diagnostic Safety in Healthcare

Diagnostic errors pose a significant risk to patient safety, yet many current Large Language Model (LLM) systems approach diagnosis as a simple prediction task, often lacking robust safeguards. Researchers have introduced AegisDx, a novel safety-oriented framework designed for hypothetico-deductive clinical reasoning. This framework orchestrates specialized LLM components using role-specific contracts, structured intermediate outputs, and evidence-retrieval interfaces, incorporating verification gates to ensure comprehensive differential diagnoses. AegisDx explicitly screens for critical "must-not-miss" conditions and rigorously verifies its reasoning against grounded medical evidence, structuring actionable next steps for clinicians. Evaluations showed substantial improvements: Top-3 diagnostic accuracy increased by 7-11% on literature cases and by 17% on emergency medicine cases compared to standalone LLMs. Crucially, it identified at least one must-not-miss condition in 78% of cases, up from 52%. Blinded physician evaluations also rated AegisDx significantly safer, improving composite safety scores in real-world emergency department notes.