Multi-source medical knowledge adaptive fusion network for combinatorial medication recommendation

Combinatorial medication recommendation has emerged as a significant research direction in artificial intelligence healthcare, demonstrating transformative potential for both pharmaceutical development and clinical decision-making. While promising, its practical implementation faces multifaceted engineering challenges spanning robust data integration, scalable model deployment, and regulatory compliance. Current drug recommendation methodologies predominantly rely on modeling electronic health records (EHRs) to generate patient representations, yet critically fail to incorporate external medical knowledge to enhance recommendation decisions. This limitation results in suboptimal integration between patient-specific data and domain-specific pharmacological knowledge. To address this critical gap in clinical decision support systems, our work pioneers the synergistic fusion of structured EHR patterns with curated medical knowledge bases, thereby enhancing recommendation accuracy and clinical relevance. In this paper, we propose a medication recommendation framework based on a Multi-source medical Knowledge Adaptive Fusion (MKAF) network. The proposed framework leverages patients’ health records and diverse medical knowledge to adaptively model their intrinsic relationships, enhancing recommendation accuracy. Specifically, we first mine patients’ health records to extract patient features. Subsequently, we design a multi-source medical knowledge module that adaptively fuses patients’ health features with various medication knowledge to capture the relationship between clinical symptoms and medications, balancing the contributions of different knowledge sources for better medication recommendations. Extensive experiments conducted on two public datasets MIMIC-III and MIMIC-IV, especially compared to the previous SOTA model, F1, PRAUC, and Jaccard have improved by 1.14%, 1.45%, 1.04% and 0.51%, 1.49%, 0.74% respectively on the two datasets.