Closed-loop medical knowledge graph question answering with adaptive planning and consistency verification

Abstract

Medical knowledge graph question answering (KGQA) becomes unreliable when a question contains several coupled conditions and requires multi-hop evidence. In medical settings, failures often come from early commitment to locally plausible paths, omission of key constraints, and weak alignment between the final answer and the cited evidence. We propose a closed-loop framework for medical KGQA that combines subgoal decomposition, adaptive graph planning, unified memory, and backward consistency verification. A question is decomposed into disease, symptom, examination, treatment, and contraindication subgoals. During search, candidate expansions are ranked by relation relevance, newly covered subgoals, historical usefulness, conflict risk, and path length. The system records explored branches, unresolved conditions, conflicts, and explanation evidence in a unified memory, and uses backward checking to decide whether a branch should be accepted, locally revised, or rolled back. Experiments on GenMedGPT-5k, CMCQA, and ExplainCPE show consistent improvements in Accuracy, F1, and Faithfulness, with the clearest gains under noisy or conflicting evidence.