Conditioned Latent State-Space Network for Intervention-Aware ICU Mortality Prediction Using Longitudinal Physiological Data

Abstract

Early prediction of intensive care unit (ICU) mortality requires models capable of capturing longitudinal physiological dynamics and therapeutic interventions. This study proposes a Conditioned Latent State-Space Network (CLSSN) that integrates temporal representation learning, intervention-aware latent state transitions, probabilistic uncertainty modeling, and counterfactual trajectory generation within a unified deep learning framework. The model was evaluated on a curated cohort of 2,000 adult ICU patients from the MIMIC-III database using six physiological variables and vasopressor administration. Experimental results demonstrated competitive predictive performance, achieving an AUROC of 0.7911, AUPRC of 0.5209, Expected Calibration Error of 0.0722, and Brier Score of 0.0973. Additional analyses, including calibration assessment, ablation studies, temporal saliency mapping, and decision curve analysis, demonstrated the framework's reliability, interpretability, and potential utility for AI-assisted critical care decision support.