Development and validation of an interpretable machine-learning-based model incorporating time-weighted variables for predicting intensive care unit mortality in cardiac arrest patients: A retrospective study

Background Cardiac arrest (CA) remains a critical medical emergency characterized by high mortality rates despite advancements in resuscitation and intensive care. The heterogeneous and evolving clinical course of CA patients hinders early and accurate prognostication. This study aimed to construct and externally validate a predictive model for 28-day intensive care unit (ICU) mortality in CA patients by leveraging time-weighted clinical data and traditional as well as machine-learning approaches including Extreme Gradient Boosting (XGBoost), Gradient Boosting with Component-wise Linear Models (Glmboost), random forest (RF), Generalized Linear Models with Elastic Net Regularization (Glmnet), and multivariate Cox regression. Methods We retrospectively analyzed adult patients with CA from three large-scale ICU databases: eICU Collaborative Research Database (eICU-CRD) (training cohort, n =1,077), Medical Information Mart for Intensive Care IV (MIMIC-IV v3.1), and MIMIC-III Clinical Database CareVue subset v1.4 (combined validation cohort, n =368). Time-weighted averages (TWA) of clinical parameters—including the Glasgow Coma Scale (GCS), vital signs, and arterial blood gas indices—were computed over the first three ICU days. Variable selection was performed using the least absolute shrinkage and selection operator (LASSO) regression, followed by multivariable Cox regression analysis. Five prognostic models—Glmboost, XGBoost, RF, Glmnet, and Cox regression—were developed and evaluated using C-index, time-dependent area under the curve (AUC), Graf score, and decision curve analysis (DCA). Model interpretability and time-varying effects were examined via SHapley Additive exPlanations (SHAP) and the Piece-wise exponential additive mixed model (PAMM). Results Nine independent predictors of ICU mortality were identified, with time-weighted average Glasgow Coma Scale (TWA-GCS) (hazard ratio: 0.75, 95% CI: 0.71–0.78, P<0.01) and duration of mechanical ventilation (hazard ratio: 0.72, 95% CI: 0.69–0.75, P<0.01) showing the strongest associations. In external validation, the Glmboost model showed the lowest Graf score. SHAP analyses underscored TWA-GCS and duration of mechanical ventilation as the most impactful variables. PAMM analysis revealed an increasingly protective temporal effect of higher TWA-GCS values. Subgroup analysis showed 28-day ICU mortality rates of 76%, 57%, and 32% in patients with low (<4), moderate (4–10), and high (>10) TWA-GCS scores, respectively (P<0.01). Conclusion Incorporating time-weighted clinical data significantly improves ICU mortality prediction in cardiac arrest patients. Glmboost showed the best calibration/error-related performance based on Graf score while demonstrating comparable discrimination, with TWA-GCS identified as a key prognostic factor.