Prototype-Aligned Multi-View Graph Learning for EHR Prediction

Electronic health records (EHRs) are a valuable resource for clinical prediction because they capture rich longitudinal information about patient care. Existing graph-based EHR models often learn relations at the cohort level, but do not clearly distinguish between global event associations and the subset of interactions actually instantiated within an individual visit. To address this limitation, we propose MVP-EHR, a multi-view prototype learning framework for EHR prediction with shared type-aware prototype alignment. MVP-EHR combines three key ideas: (i) a lift-thresholded global knowledge graph that captures cohort-level event associations, (ii) a visit-induced local knowledge graph that preserves only the observed events and supported event-event relations within each visit, and (iii) shared type-aware prototypes that align local and global event representations before post-event multi-view fusion and temporal prediction. Experiments on MIMIC-IV show that MVP-EHR achieves strong and consistent performance across multiple clinical prediction tasks, including 0.9339 AUROC and 0.1706 AUPRC for mortality prediction, the best AUROC on readmission prediction, the best overall performance on phenotype prediction, and a strong AUPRC/F1 tradeoff on drug recommendation among the compared baselines. Ablation results further show that separating global prior from visit-instantiated evidence, together with shared prototype alignment, is critical for robust performance.