Improving mortality prediction in critically ill cancer patients with a multidimensional machine learning model

BACKGROUND: Prognostic assessment in critically ill cancer patients is challenging due to the suboptimal performance of traditional severity scores. We developed and validated machine learning models to provide an objective triage and risk-stratification tool at admission, identifying patients who require high-intensity organ support. METHODS: We conducted a retrospective cohort study including 997 critically ill cancer patients admitted to the ICU. Forty-six demographic, oncologic, physiological, laboratory, and therapeutic variables collected at ICU admission were used to train and validate ML models. Eight algorithms were evaluated using stratified cross-validation with feature selection and hyperparameter optimization. Model performance was assessed using discrimination, calibration, and classification metrics. Model interpretability was explored using Shapley additive explanations (SHAP). RESULTS: CatBoost achieved the best performance for ICU mortality prediction (AUROC 0.94), showing excellent discrimination and calibration, and outperforming other ML models. Prediction of 30-day survival was less accurate (best AUROC 0.74), reflecting the influence of post-ICU factors not captured at admission. Key predictors of ICU mortality included severity of organ dysfunction, therapeutic objectives, vasopressor and methylene blue use, SAPS III score, lactate, platelet count, and blood urea nitrogen. For 30-day survival, baseline physiological status, admission type, SAPS III, lactate, creatinine, age, and body mass index were most relevant. SHAP analysis demonstrated that acute physiology and organ dysfunction, rather than cancer diagnosis alone, primarily drove short-term outcomes. CONCLUSIONS: ML-based models, particularly CatBoost, outperformed traditional tools in predicting ICU mortality. Within this oncologic cohort, short-term outcomes were driven primarily by acute physiological derangements rather than specific cancer characteristics. These results support using ML-based risk stratification as an objective triage tool at admission. External validation is warranted to confirm generalizability before clinical integration.