Critically ill patients with ischemic stroke face substantial in-hospital mortality. Early and accurate prediction of mortality risk may facilitate timely risk stratification and improve intensive care management. This study aimed to develop and externally validate a machine learning–based predictive model using multicenter datasets. We extracted data from MIMIC-IV (n = 3,568), eICU-CRD (n = 2,535), and Tongji University Hospital (TJUH, n = 144). Eight predictors identified through multivariable logistic regression were used for model construction. Five algorithms—Random Forest, XGBoost, LightGBM, Logistic Regression, and an ensemble SuperLearner—were trained and evaluated. Model performance was assessed using the area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, F1 score, and calibration. DeLong’s test compared the AUROC of the SuperLearner with those of the four base models. Decision curve analysis (DCA) was used to evaluate clinical utility. The SuperLearner demonstrated the strongest discrimination, with AUROCs of 0.80 (95% CI: 0.75–0.85) in MIMIC-IV, 0.82 (95% CI: 0.80–0.84) in eICU-CRD, and 0.76 (95% CI: 0.64–0.87) in TJUH. Sensitivity and specificity ranged from 0.70 to 0.83 and 0.68–0.82 across datasets, outperforming the four baseline models (all p < 0.05). Calibration curves showed reasonable agreement between predicted and observed outcomes, although deviations were noted at probability extremes. DCA indicated net clinical benefit across a wide range of thresholds, while slight fluctuations suggest potential overestimation at higher thresholds. The SuperLearner model achieved robust and consistent predictive performance across three independent cohorts. While showing potential utility for early mortality risk estimation in critically ill ischemic stroke patients, its clinical application requires further prospective validation and real-world implementation studies.
Hu et al. (2026) studied this question.