Collaborative Hypergraph Networks for Enhanced Disease Risk Assessment

Electronic Health Records (EHRs) are a treasure trove of patient data, pivotal for predicting medical incidents. Despite advances in deep learning, challenges persist in leveraging EHRs for effective disease prediction. This paper presents a novel approach, the Collaborative Hypergraph Network-based Disease Prediction method (CHLN), which addresses these challenges by capturing the nuanced interplay between patients and diseases within EHRs. Our method categorizes diagnosed diseases into chronic and acute, constructs a dynamic hypergraph to model high-order disease relationships, and enriches the clinical event representation through medical language model encoding. By integrating these representations, we formulate a comprehensive patient representation that predicts future diagnoses. We rectify the neglect of horizontal disease relationships and enhance patient similarity feature capture through an attention network that learns disease code hierarchies. Experiments on benchmark datasets reveal that CHLN significantly surpasses existing models, underscoring the efficacy of our collaborative hypergraph framework in medical informatics.