StructCare: Dynamic graph structure learning for enhancing context-aware healthcare

Abstract

The advancement of intelligent healthcare systems has generated large-scale electronic health records (EHRs). This data provides a foundation for advancing machine learning in healthcare, particularly for pivotal tasks like disease prediction and medication recommendation. Despite the growing focus on modeling relationships within EHR, most existing methods treat medical events as connected by static relationships, where associations are defined by co-occurrence or fixed correlations. However, such approaches overlook dynamic relationships that reflect how a patient’s health status and treatment evolve over time. Therefore, we propose StructCare, a clustering-based graph structure learning framework that explicitly models these dynamic relationships by integrating lab test results with medical events. Specifically, StructCare leverages large language models to construct relationships among medical events and generate personalized patient graphs from visit records. It then employs temporal networks to capture evolving health trends from monitoring-level events and dynamically updates the graph structure at each visit, generating richer patient representations, thereby improving the accuracy of disease prediction and medication recommendation. Extensive experiments on two real-world datasets and across two tasks show that StructCare consistently outperforms state-of-the-art methods. Specifically, it achieves an average improvement of approximately 2.1 % in F1-score and 2.8 % in Jaccard for disease prediction, and 2.3 % in F1-score and 4.0 % in Jaccard for medication recommendation, highlighting its effectiveness in capturing dynamic, context-aware relationships.