Development and validation of a machine learning-based model for predicting delirium risk in postoperative brain tumor patients in the Intensive Care Unit

Objective This study aimed to clarify the incidence and influencing factors of delirium in ICU patients after brain tumor surgery, construct and validate delirium risk prediction models using multiple machine learning algorithms, identify the optimal model, and develop a personalized risk calculation tool to provide evidence-based support for early precise identification of high-risk patients and targeted preventive interventions. Methods Consecutive convenient sampling was adopted. A total of 600 patients who underwent brain tumor surgery in a Grade A tertiary hospital in Nanchang (July 2021–December 2024) served as the modeling cohort, and 160 similar patients (January–August 2025) as the external validation cohort. LASSO regression screened independent risk factors from 26 candidates. Five models (LR, ANN, RF, DT, SVM) were constructed and evaluated by sensitivity, AUC, DCA, etc. SHAP interpreted the optimal model’s feature importance. Results Five independent risk factors were identified: age, APACHE II score, postoperative GCS score, use of diuretics or dehydrating agents, and mechanical ventilation duration. The SVM model performed best: validation set AUC = 0.857 (95%CI:0.814–0.898), test set AUC = 0.847, external validation accuracy 70.00%. SHAP showed diuretics or dehydrating agents were the most important feature (mean |SHAP|=0.1625). An online risk calculator was developed for convenient personalized assessment. Conclusions The SVM-based predictive model has excellent efficacy and generalizability. The easy-to-operate risk calculator can effectively identify high-risk patients early, providing scientific support for precise delirium prevention and control.