KnowAugNet: Multi-Source Medical Knowledge Augmented Medication Prediction Network with Multi-Level Graph Contrastive Learning
Predicting medications is a crucial task in many intelligent healthcare systems. It can assist doctors in making informed medication decisions for patients according to electronic medical records (EMRs). However, medication prediction is a challenging data mining task due to the complex relations between medical codes. Most existing studies focus on utilizing inherent relations between homogeneous codes of medical ontology graph to enhance their representations using supervised methods, and few studies pay attention to the valuable relations between heterogeneous or homogeneous medical codes from history EMRs, which further limits the prediction performance and application scenarios. Therefore, to address these limitations, this paper proposes KnowAugNet, a multi-sourced medical knowledge augmented medication prediction network which can fully capture the diverse relations between medical codes via multi-level graph contrastive learning framework. Specifically, KnowAugNet first leverages the graph contrastive learning using graph attention network as the encoder to capture the implicit relations between homogeneous medical codes from the medical ontology graph and obtains the knowledge augmented medical codes embedding vectors. Then, it utilizes the graph contrastive learning using a weighted graph convolutional network as the encoder to capture the correlative relations between homogeneous or heterogeneous medical codes from the constructed medical prior relation graph and obtains the relation augmented medical codes embedding vectors. Finally, the augmented medical codes embedding vectors and the supervised medical codes embedding vectors are retrieved and input to the sequential learning network to capture the temporal relations of medical codes and predict medications for patients.