Temporal Modeling of Longitudinal Patient Trajectories Using Recurrent Neural Architectures for Prognostic Risk Stratification in Multicenter Clinical Cohorts

Abstract

Prognostic risk stratification is crucial for optimizing care delivery in clinical settings. This paper investigates the application of recurrent neural network (RNN) architectures, such as LSTM and GRU, for modeling longitudinal patient trajectories across multicenter datasets. We propose a temporal deep learning framework that captures time-varying dependencies and nonlinear patterns in multivariate electronic health records (EHRs). The model demonstrates significant performance gains in predicting adverse outcomes such as mortality and readmission, compared to traditional models. Using datasets from diverse clinical cohorts, our approach achieves up to 20% improvement in AUC-ROC for early prediction of patient deterioration. This study underscores the potential of recurrent architectures for clinical decision support in real-world settings.