Longitudinal prediction from electronic health records (EHR) suffers from sparse and irregular patient trajectories, and external knowledge graphs (KG) offer a way to mitigate these issues. Most existing approaches use a fixed, context‑agnostic topology augmentation that adds the same KG nodes and edges regardless of a patient’s evolving state.
We introduce ReTA, a reinforcement‑learning‑based dynamic topology augmentation framework that treats KG import as a per‑visit, budget‑aware policy. ReTA first builds an offline refined pool of KG‑grounded templates, then learns a policy to pick one action per visit from three options: Soft Import, which enriches node features without altering graph topology; Hard Import, which grafts a compact KG subgraph onto the visit graph to create message‑passing shortcuts; and Skip, which leaves the visit unchanged when the base encoder is already confident.
To stabilize learning, ReTA employs a decoupled encoder that processes semantic and structural signals in separate channels and fuses them via adaptive gating.
Experiments on MIMIC‑III and MIMIC‑IV for diagnosis prediction, mortality, and readmission show that ReTA consistently outperforms strong baselines while remaining efficient. It transfers across datasets and knowledge graphs and yields interpretable augmentation patterns. The robust gains under sparse supervision highlight the advantage of ReTA’s dynamic decision to import knowledge, boosting accuracy while curbing costs.
Review: ReTA leverages reinforcement learning to adaptively import KG knowledge according to patient state, overcoming the limitations of static augmentation and delivering interpretable, cost‑effective improvements on real clinical data.