Relation-aware dynamic knowledge subgraph enhanced large language models for medical multiple-choice question answering

Abstract

Medical multiple-choice question answering requires models to simultaneously understand question semantics, candidate-answer differences, and structured relations among medical entities in complex clinical contexts. To address the limitations that existing knowledge-enhanced methods often adopt question-level unified modeling and thus struggle to characterize local relational differences among different candidate options, while conventional low-rank adaptation lacks relation-pattern adaptivity, this paper proposes a knowledge-enhanced large language model framework that combines option-level knowledge subgraph representation fusion with relation-aware dynamic LoRA adaptation. The proposed method first constructs a local medical knowledge subgraph for the question stem and each candidate answer separately, and then injects structured knowledge into candidate representations through a graph encoder and gated fusion mechanism. Subsequently, multiple learned LoRA expert branches are dynamically combined through candidate-conditioned routing based on the relational information encoded in the enhanced candidate representations, enabling the model to obtain more suitable parameter update paths. The experimental results are reported as the mean and standard deviation over three random-seed runs. On the MedQA dataset, the proposed method achieves Accuracy scores of 62.90 % ± 0.82 % and 62.84 % ± 1.04 % , together with F1-Score values of 62.97 % ± 1.07 % and 62.79 % ± 0.78 % , using Llama3 8B and Qwen 7B, respectively. On the MedMCQA dataset, the corresponding Accuracy scores are 61.04 % ± 0.30 % and 58.59 % ± 0.65 % , while the F1-Score values reach 61.21 % ± 0.54 % and 57.77 % ± 0.57 % , respectively. Experimental results show that the proposed method can effectively improve knowledge understanding, candidate discrimination, and reasoning stability in medical multiple-choice question answering.