ReTA Framework Dynamically Augments EHR Data with External Knowledge

Longitudinal prediction using Electronic Health Records (EHRs) is often hampered by the inherent sparsity and irregularity of patient data. Augmenting EHR graphs with external knowledge graphs (KGs) offers a promising solution, but existing methods typically apply fixed, context-agnostic augmentations, adding the same KG nodes regardless of a patient's evolving condition. This research introduces ReTA (Reinforcement learning-based dynamic Topology Augmentation), a novel framework that treats KG import as a per-visit, budget-aware policy decision. ReTA first creates a refined pool of KG-grounded templates offline. Then, it learns a policy to select one of three augmentation actions for each patient visit: "Soft Import" (enriching node features), "Hard Import" (grafting a compact KG subgraph), or "Skip" (leaving the visit unaugmented if the base encoder is already confident). To ensure stable learning, ReTA employs a decoupled encoder that processes semantic and structural signals separately before fusing them. Experiments on MIMIC-III and MIMIC-IV datasets for tasks like diagnosis prediction, mortality, and readmission consistently show ReTA outperforming strong baselines. It remains efficient, transfers across datasets and KGs, and provides interpretable augmentation patterns, highlighting the benefits of dynamic, context-aware knowledge integration, especially under sparse supervision.