A Hybrid Transformer–Graph Neural Network Architecture for Personalized Clinical Risk Prediction Using Electronic Health Records

Abstract

Electronic health records (EHRs) contain longitudinal measurements, diagnoses, procedures and medication events that jointly encode clinical deterioration. Sequence models can represent temporal progression but may overlook explicit relations among clinical concepts, whereas graph neural networks (GNNs) capture relational structure but often compress chronology. This manuscript proposes a hybrid Transformer–GNN architecture for personalized prediction of adverse clinical outcomes at a predefined horizon. A causal Transformer processes timestamped patient events; a patient-specific heterogeneous graph represents diagnoses, laboratory findings, medications and encounters; and a gated fusion mechanism combines temporal and relational representations. We specify cohort construction, leakage prevention, missingness modeling, training objectives, calibration, subgroup auditing and prospective evaluation. A reproducible evaluation protocol is described for a credentialed EHR resource such as MIMIC-IV. Because no patient-level experiment was performed for this manuscript, numerical results are explicitly labeled as synthetic demonstration values, not clinical validation or evidence of superiority. The contribution is a detailed and testable design intended for future external validation.