Early prediction of prolonged ICU stay in sepsis patients using an explainable hybrid deep learning model

Abstract Sepsis remains a leading cause of intensive care unit (ICU) admission and is frequently associated with prolonged ICU stay, which contributes to increased morbidity, healthcare utilization, and strain on critical care resources. Despite widespread use, existing clinical severity scores inadequately capture the individual risk of prolonged ICU stay in sepsis. We developed a hybrid deep learning (FT-TabNet) model with an adaptive gated fusion framework to predict prolonged ICU stay (>4 days) using 17 routinely collected Sequential Organ Failure Assessment (SOFA)-based features available within the first 24 hours of ICU admission. Model development and internal validation were performed using the MIMIC-IV dataset, with external validation conducted in a single-center cohort from Chungbuk National University Hospital (CBNUH) and Chungnam National University Hospital (CNUH), South Korea. The hybrid model showed strong discriminative performance in internal validation, achieving an AUROC of 0.848, and consistent performance across the external validation cohort, with modest variability, achieving an AUROC of 0.823 in the CBNUH and 0.781 in the CNUH cohort. SHAP-based explainability enabled transparent, patient-level interpretation of key contributors to prolonged ICU stay risk. These findings suggest that an explainable hybrid deep learning approach can help clinicians to identify patients at risk of prolonged ICU stay early in sepsis, using routine clinical data to support clinical decision-making and ICU resource allocation.