AutoClinician: Structured Clinical Guideline Integration for Trustworthy Diagnostic Reasoning in Healthcare

ICLR 2026 Conference Submission15208 Authors

Keywords: healthcare, LLMs, Conversational System, Diagnosis Prediction, Clinical Guideline

Abstract: Recent advances in large language models (LLMs) have enabled clinical conversational systems with impressive diagnostic capabilities. However, existing approaches often lack alignment with real-world clinical workflows and fail to provide interpretable, evidence-grounded reasoning. In this work, we propose AutoClinician, a unified and training-free framework that integrates clinical guidelines to support stepwise and explainable diagnosis on real-world electronic health records (EHRs). AutoClinician first extracts and summarizes narrative guidelines into Clinical Evidence Graphs (CEGs). These graphs are further automatically verified and refined using a consistency-based strategy. To support trustworthy and patient-specific diagnosis, we utilize CEGs by conducting context-aware, evidence-grounded clinical reasoning on EHRs with Deterministic Finite Automaton (DFA). Our framework outperforms both general-purpose and clinically specialized LLMs, and exhibits stronger interpretability.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 15208