DxChain: A Cognitive AI Agent for Clinical Diagnosis via LLMs

Researchers propose DxChain, a novel chain-based clinical reasoning framework that uses large language models (LLMs) to improve clinical diagnosis from unstructured electronic health records (EHRs). DxChain addresses challenges of 'tunnel vision' and diagnostic hallucinations by mirroring a clinician's cognitive trajectory through three phases: Memory Anchoring, Navigation, and Verification. Key innovations include a Profile-Then-Plan paradigm to establish a panoramic patient baseline, a Medical Tree-of-Thoughts (Med-ToT) algorithm for strategic planning, and a Dialectical Diagnostic Verification procedure using 'Angel-Devil' adversarial debates. The framework aims to mitigate cold-start hallucinations and enhance diagnostic accuracy.