ICONQUER: A Transformer-Based Instruction-Finetuned Context-Aware Medical Question Answering Model With Knowledge Graph Augmentation

Abstract

Medical Question Answering (MQA) systems are vital for clinical decision support, yet most existing models struggle with complex query processing, fine-grained context awareness, and explainability. This study introduces ICONQUER, a transformer-based MQA model that integrates instruction-finetuned embeddings and knowledge graph (KG) augmentation to improve semantic coherence, contextual relevance, interpretability and response accuracy. We designed five sequential experiments to evaluate ICONQUER across MedQA and HotPotQA datasets: (i) representation analysis of multiple embeddings, (ii) answer generation on MedQA development data, (iii) validation on the MedQA test set, (iv) robustness testing with external biomedical knowledge sources (Wikipedia, BioPortal), and (v) cross-domain generalization on HotPotQA. Results show that ICONQUER achieves the strongest semantic alignment (cosine similarity = 0.9446 on MedQA train; 0.9387 on MedQA test; 0.914 on HotPotQA), consistently outperforming state-of-the-art MQA systems leveraging E5, BioBERT, and BERT-Large baselines. While external knowledge enrichment yielded limited benefits, knowledge graph integration substantially improved interpretability and the capture of semantic relationships. This advancement enabled more contextually precise answers, positioning ICONQUER as a promising tool for clinical decision-making. Future work will extend ICONQUER by incorporating broader biomedical ontologies (e.g., SNOMED CT, UMLS), clinical trial repositories, and electronic health records. These enhancements aim to improve adaptability across medical subfields and to strengthen explainability through intuitive medical reasoning, thereby supporting global applicability. Further efforts will address the integration of advanced AI techniques, reinforcement of data security, support for multilingual use, and improvements in transparency and flexibility.