Beyond EHRs: External Clinical knowledge and cohort Features for medication recommendation

Medication recommendation plays an important role in healthcare by supporting clinical decision-making, while rich clinical experiences and knowledge are key factors that contribute to the success of this task. However, existing methods still face certain limitations: the acquisition of clinical experience is confined to the historical records of an individual patient, and encoded clinical knowledge is typically used as coarse auxiliary information to enhance encoded records. To address these limitations, we propose the EXternal Clinical knowlEdge and cohoRt Features (EXCERF) model for the recommendation task. EXCERF constructs an external Memory Neural Network shared across all patients to capture cohort features tailored to specific patient groups, allowing patients to access relevant clinical experiences distributed across a large cohort. Then the model incorporates clinical knowledge tuples as additional biases to refine interactions between clinically related entities within the self-attention mechanism, concurrently considering detailed clinical and semantic knowledge to build patient representations. Finally, EXCERF integrates representations of multiple admissions with GRU to generate the final recommendation. Experimental results on real-world clinical records demonstrates that EXCERF achieves superior performance and facilitates effective medication recommendation.