Objective To develop and validate a prediction model for in-hospital recurrent intensive care unit (ICU) admission in critically ill patients with ischemic stroke (IS) based on machine learning (ML) algorithms.

Methods Clinical data from 2 929 IS patients were included from the Medical Information Mart for Intensive Care IV (MIMIC-IV) database. Least absolute shrinkage and selection operator (LASSO) regression was used to identify predictive factors, and the synthetic minority over-sampling technique (SMOTE) was employed to create a derivation cohort comprising 2 583 patients. These patients were randomly divided into a training set (n=2 066) and a test set (n=517) at an 8:2 ratio. Five ML algorithms, including decision tree, random forest, adaptive boosting (AdaBoost), gradient boosting decision tree (GBDT), and support vector machine (SVM), were performed to construct prediction models. Five-fold cross-validation was used to evaluate the performance of the model in the training set. The area under the receiver operating characteristic curve (ROC-AUC) and decision curve analysis (DCA) were used to assess and compare the models in the testing set. The best-performing model was interpreted by shapley additive explanations (SHAP).

Results Among the 2 929 patients included, 704 (24.0%) experienced in-hospital recurrent ICU admission. Among the five ML models, the random forest model demonstrated the best predictive performance, with an AUC of 0.839 (95%CI 0.801–0.877). Feature importance analysis identified five most significant features affecting model prediction, including APS Ⅲ score, albumin, age, heart rate, and SOFA score.

Conclusions ML-based models can effectively predict the risk of in-hospital recurrent ICU admission in critically ill patients with IS. The random forest model showed superior predictive performance, which may have potential applications in early clinical risk stratification and intervention.