A Multi-Graph Fusion Framework for Patient Representation Learning

Deep learning has increasingly been used to model electronic health records (EHR) data for a wide range of medical analysis, such as clinical risk prediction. Existing methods have focused on patient representation learning from a single graph view. In real clinical reasoning scenarios, it is a common practice to integrate information from different patient-level features (e.g., demographics, vital signs, diagnoses, procedures, and lab tests) to derive a patient health context, which can naturally be mapped to deep learning with multiple graphs generated from the patient-level features. Confronting the challenge of learning patient representations in clinical risk prediction, we present a new Multi-Graph Fusion Framework for patient representation learning, which learns multiple graph structures from input patient-level features and, in turn, generates an optimal graph structure that incorporates the characteristics of these graphs with attention mechanisms. Our method further aggregates the information from similar patients to offer a rich representation of the patient, which allows extraction of patient health context for missing data imputation and clinical risk prediction. Evaluation using two real-world EHR databases demonstrates the effectiveness and superiority of our method over competitive baselines.