Drug Recommendation Combining Attention Mechanism and Graph Encoder

Drug recommendation is an vital mission in the field of intelligent healthcare. Since patients often suffer from multiple diseases and the disease status is dynamic, the model must take drug interactions into account when recommending drugs. However, most existing drug recommendation works only rely on patient electronic health records for medical predictions, current methods do not fully exploit the patient information and their interrelationships when constructing patient narratives, resulting in insufficient completeness of patient narratives. They also overlook the future trend of patient's condition changes, resulting in a high rate of DDI in recommended drug combinations. To work around these limitations, this paper proposes a drug recommendation model that combines attention mechanism and graph encoder. Encode different types of medical codes in patient presentations using access level codes and attention systems. At the same time, a drug graph encoder is designed to extract and encode key information based on the structural features of drug molecules. Comparison with current state-of-the-art basic model, the proposed model achieves a large degree of improvement.