Temporal Graph Attention Model for Enhanced Clinical Risk Prediction

Electronic health record (EHR)-based clinical risk prediction can help clinicians make better decisions and understand early diagnosis. Nevertheless, accurate representations derived from multi-dimensional time-series electronic health record data are crucial to the prediction performance. Most current systems either extract data in two stages, concentrate on temporal aspects, or assume that clinical event variables have intrinsic links. Prediction accuracy suffers because of a lack of patient feature information. Additionally, selecting the appropriate Meta-Paths is often a manual process in most current systems that rely on Heterogeneous Graph Neural Networks. Our proposed Time-aware Context-Gated Graph Attention Network aims to address these issues (TContext GAN). To automatically choose Meta-Paths and extract temporal semantic information and inherent relations from EHR data, we developed a GNN-based module with a selfattention mechanism and Time-aware Meta-Paths.