AI-Powered Predictive Model for Early Detection of Disease Progression using Patient Health Data

Abstract

Routine clinical metrics may miss subtle physiological variations that occur before clinically evident disease progression, delaying preventive intervention. This study presents an AI-powered predictive model trained on 4,567 de-identified longitudinal patient records containing 36 structured clinical variables, 14 laboratory biomarkers, demographic descriptors, medication history, and three years of follow-up. Disease progression was labelled using a predefined composite endpoint combining sustained biomarker deterioration, clinically documented worsening, treatment escalation, or disease-related hospitalization within the follow-up period. Missing values were managed using temporally constrained forward-backward interpolation with missingness indicators, class imbalance was examined using progression and non-progression distributions, and privacy was maintained through de-identification and controlled data handling. A hybrid temporal convolutional encoder and gated recurrent prediction module captured short-term fluctuations and long-range trends. The model was trained on 3,214 records and tested on 1,353 records, achieving 94.27% accuracy, 0.962 AUC, and 0.943 F1-score, while reducing false negatives compared with conventional clinical scoring and sequential baselines. Generalization was assessed through patient-level hold-out testing, temporally separated validation, and stratified five-fold cross-validation because a fully independent external cohort was not available. These findings support the potential of the model for early-warning prediction and clinically timely intervention.