Interaction-Aware Temporal Prescription Generation via Message Passing Neural Network

In recent years, deep learning techniques have been applied with great success in the healthcare industry, such as disease prediction and drug recommendation. However, existing works on drug recommendation either do not take the critical impact of doses on treatment outcomes into account, or neglect the patient’s personalized history and drug-drug interactions, resulting in suboptimal recommendation results. To fill these gaps, we propose a novel Temporal Prescription Generation (TPG) model in this paper. Specifically, we first utilize a message passing neural network to capture the interaction between drugs, combined with a decomposed long and short-term memory (LSTM-DE) to characterize patient’s general information. Then, we design a prescription generator based on recurrent neural network to generate drugs and estimate doses simultaneously. Experimental results on a real-world dataset clearly demonstrate that the proposed model is superior to the best existing models in terms of drug recommendation and dose estimation.