PROBLEM: Current LLM-based CDSS only provide post-hoc explanations, not live transparency. Clinicians cannot observe diagnostic reasoning as it unfolds. Multi-agent reasoning generates long, overwh..
PROBLEM: Current LLM-based CDSS only provide post-hoc explanations, not live transparency. Clinicians cannot observe diagnostic reasoning as it unfolds. Multi-agent reasoning generates long, overwhelming token streams. GAP: No existing CDSS supports token-level, real-time explainability. Flow control and semantic interpretation are tightly coupled in current systems. Prior work focuses on static summaries or rule-based rationales. PURPOSE: Objective: Design EXAID, a middleware that summarizes evolving diagnostic reasoning as tokens are streamed. Goal: Enable clinicians to monitor live logic without being overwhelmed by raw data.
