Structural and Temporal Relationship Modelling of Electronic Health Records for Medical Risk Prediction

Analysis of historical electronic health record (EHR) data for predicting the medical risks of patients is an essential task in personalized healthcare. Each patient has a sequence of hospital visits, and each visit contains an unordered set of medical codes. We identify two major aspects of EHRs: (1) structural relationships between visits and codes (i.e., code-sharing behaviours of visits) represented with a graph data structure, and (2) temporal relationships in the visit sequences (i.e., disease progression) modelled by a point process. We propose novel deep learning models to capture both these relationship types for accurate and interpretable medical risk prediction.