Explainable machine learning prediction of tracheostomy after craniotomy for supratentorial intracerebral hemorrhage

Accurate prediction of tracheostomy after craniotomy for supratentorial intracerebral hemorrhage (sICH) remains challenging. This study aimed to develop, externally validate, and interpret a machine learning model for individualized risk prediction. A retrospective multicenter cohort was constructed, including 738 patients from Weifang People's Hospital and 186 from Weifang Hospital of Traditional Chinese Medicine who underwent craniotomy between January 2017 and December 2024. Predictor variables were screened using least absolute shrinkage and selection operator (LASSO) and multivariate logistic regression. Logistic regression, random forest, and extreme gradient boosting (XGBoost) models were trained with repeated 10-fold cross-validation and assessed for discrimination, calibration, and clinical utility. Five key predictors were identified: Glasgow Coma Scale, age, hematoma volume, operative time, and serum bicarbonate. In external validation, XGBoost demonstrated the most balanced and robust performance, with an AUROC of 0.86 and a Brier score of 0.15, and showed superior net benefit on decision curve analysis. SHapley Additive exPlanations confirmed clinical plausibility, and a web-based dynamic nomogram was developed for individualized prediction. This explainable XGBoost model provides reliable and interpretable estimation of postoperative tracheostomy risk, facilitating evidence-based perioperative decision-making and resource allocation in neurocritical care.