Graph Neural Networks for Disease Progression Prediction in Healthcare Networks

Abstract

Predicting how a patient's condition will evolve is central to preventive and personalized medicine, yet most predictive models treat each patient as an isolated sequence and ignore the rich relational structure that connects patients, visits, diagnoses, medications, and laboratory phenotypes within a healthcare network. Recurrent and Transformer models capture temporal patterns but discard inter-entity dependencies, while static graph neural networks (GNNs) capture relations but not the time-ordered nature of disease progression. This paper proposes ProGNet, a relation-aware heterogeneous temporal graph attention network that represents longitudinal electronic health records (EHR) as a typed healthcare graph and fuses structural and temporal learning in a single end-to-end framework. ProGNet performs type-specific multi-head attention over patient, visit, diagnosis, medication, and lab nodes, pools per-visit graph embeddings, and feeds them through a timeaware encoder with explicit inter-visit gap decay to model progression. A progression head predicts future disease stage and horizon-specific risk, an auxiliary comorbidity link-prediction task regularizes representation learning, and attention plus sub-graph explanations provide transparency. Evaluated on longitudinal EHR cohorts with patient-level splitting and five-fold cross-validation, ProGNet attained 94.3% AUROC, 91.2% AUPRC, a 90.4% macro-F1, and a 0.897 concordance index, outperforming the strongest baseline by 2.7 AUROC points and degrading more slowly as the prediction horizon lengthens. The principal contributions are: (i) a unified heterogeneous-temporal graph formulation for progression; (ii) a relation-aware attention mechanism with time-gap decay; and (iii) an explainable, ablation-validated pipeline. Results show that jointly modeling network structure and temporal dynamics, rather than either alone, drives accurate and robust progression prediction