Diagnostic consultation is an online sequential decision‑making process where clinicians gather evidence through patient interaction until a diagnosis is sufficiently supported. Automating this process requires adaptive questioning and interpretable decisions. Bayesian networks naturally update diagnostic posteriors as evidence accumulates, but open‑ended consultation raises two challenges: linking diagnostic hypotheses to potential inquiries and converting evolving posteriors into consultation actions.
We introduce AutoDisym, an automated pipeline that leverages heterogeneous diagnosis‑labeled clinical narratives to construct a Disorder–Symptom Bayesian Network (DSBN). AutoDisym extracts disease‑symptom pairs, then uses GPT‑5.6‑Sol for structure learning and parameter estimation, automatically producing high‑quality DSBNs across psychiatry, respiratory medicine, fever clinics and three public datasets.
Built on the DSBN, ConsultMind is an uncertainty‑aware framework that updates disorder posteriors after each patient response and uses posterior uncertainty (e.g., entropy) to guide the next inquiry or to issue a diagnosis. Experiments show AutoDisym achieves a macro‑averaged F1 of 81.37 % for canonical symptoms and 72.19 % for manifestations. ConsultMind improves Top‑1 accuracy by up to 22.15 percentage points and Top‑3 by up to 37.89 points. Physician evaluation confirms that ConsultMind yields better ranking explanations, differential diagnoses, and rationales across LLMs of various scales.
This study demonstrates a promising approach to automatic, interpretable, and uncertainty‑aware diagnostic consultation by integrating domain knowledge with large language models.
Review