Graph Attention Network and Dynamic Adjustment Mechanism for Drug Recommendation

Drug recommendation is an important part of healthcare. Leveraging electronic health records for drug recommendation can assist doctors to make better decisions. While deep learning has made some progress in drug recommendation, further efforts are needed to improve their accuracy and safety. The research of drug combination recommendation involves critical works such as representation of drug combination, safety of drug combination and so on. For representation of drug combination, traditional drug recommendation methods overlook the importance of drug co-occurrence among different historical drug combinations. This work uses graph attention network to calculate the different weight of drug combination co-occurrence. For safety of drug combination, this work designs a mechanism to dynamically adjust the recommendation strategy for balancing accuracy and security. The experimental results clearly show that this method has demonstrated significant effectiveness in improving the accuracy and safety of recommendation.