Predicting delirium in critically Ill patients with cancer using interpretable machine learning

Delirium is common in critically ill patients with cancer, but early recognition remains difficult and externally validated prediction tools are limited. We aimed to develop and externally validate interpretable machine learning models and identify clinically meaningful factors associated with delirium. We conducted a retrospective multicohort prediction study involving 20,596 adults with cancer admitted to intensive care: 5,686 patients from the Medical Information Mart for Intensive Care IV database were used for model development and internal validation, and 14,910 patients from the eICU Collaborative Research Database were used for external validation. Five machine learning models were evaluated using discrimination, calibration, and decision curve analysis. Predictor contributions were assessed using Shapley additive explanations. Here we show that the tabular foundation model provides the best internal discrimination (area under the receiver operating characteristic curve, 0.94), whereas extreme gradient boosting demonstrates the strongest external performance (0.93). Propofol exposure is the leading internal predictor, while Glasgow Coma Scale score is most influential externally. Additional predictors include antibiotic use, prolonged mechanical ventilation, greater physiological severity and organ dysfunction, potassium abnormalities, elevated bilirubin, older age, and insurance-related disadvantage. These findings indicate that baseline vulnerability and acute physiological insults jointly characterize delirium risk. Interpretable machine learning accurately identifies patients at increased delirium risk across two independent critical-care cohorts. The models may support earlier risk stratification and highlight potentially modifiable prevention targets, but prospective evaluation and clinical-impact testing are required before routine implementation. Delirium is a sudden state of confusion that often affects people with cancer who are critically ill. This study aimed to identify patients at risk early, so healthcare teams may be able to prevent or reduce its effects. We analyzed electronic health records from 20,596 adults with cancer treated in intensive care units and compared five computer-based prediction models. The models predicted delirium accurately in two large, independent patient databases. Important warning signs included older age, greater illness severity, infection, organ problems, abnormal potassium levels, longer use of breathing machines, and exposure to sedative medicines. These findings may help healthcare professionals recognize high-risk patients sooner and provide personalized monitoring and preventive care.