Accurate differential diagnosis requires well‑structured and unambiguous patient case descriptions; missing or vague information easily leads to errors and inconsistent reasoning. Large language models have shown promising performance on medical QA and clinical reasoning, yet they mainly rely on pattern recognition from training data instead of grounded clinical evidence, often producing plausible but unsupported conclusions and offering opaque reasoning that clinicians cannot verify.
To close this gap we propose an ontology‑guided multi‑agent differential diagnosis system that mimics the iterative reasoning of physicians. The system anchors inference in an expert‑curated medical ontology, representing symptom‑disease and test‑result relationships in a structured form so that each diagnostic step can be directly linked to verifiable evidence.
The architecture consists of four key components: an ontology repository supplying disease‑symptom and examination links; an agent orchestration layer that assigns tasks across reasoning stages; an evidence‑tracking engine that maps each inference to ontology relations, producing an auditable reasoning trace; and a discriminative symptom identifier that selects the most informative missing findings via information‑gain analysis and automatically generates targeted follow‑up questions to reduce uncertainty.
Evaluations demonstrate that ontology‑based constraints markedly suppress hallucinations, improve transparency, traceability and consistency, and achieve higher accuracy on public medical benchmarks. The full source code is publicly released for reproducibility and further development.
Review