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

Area of Science:

• Critical Care Medicine
• Oncology
• Data Science & Machine Learning

Background:

• Prognostic assessment for critically ill cancer patients is difficult with traditional scores.
• Machine learning (ML) models offer potential for objective risk stratification.
• Identifying patients needing intensive organ support is crucial.

Purpose of the Study:

• Develop and validate ML models for predicting outcomes in critically ill cancer patients.
• Provide an objective triage and risk-stratification tool at ICU admission.
• Compare ML model performance against traditional severity scores.

Main Methods:

• Retrospective cohort study of 997 critically ill cancer patients.
• Utilized 46 variables (demographic, oncologic, physiological, lab, therapeutic) at ICU admission.
• Trained and validated eight ML algorithms using cross-validation, feature selection, and hyperparameter optimization.

Main Results:

• CatBoost model achieved superior ICU mortality prediction (AUROC 0.94) with excellent discrimination and calibration.
• Predicting 30-day survival was less accurate (best AUROC 0.74), influenced by post-ICU factors.
• Key predictors included organ dysfunction severity, therapeutic goals, vasopressor use, SAPS III, lactate, platelets, and BUN; SHAP analysis highlighted physiology over cancer type.

Conclusions:

• ML models, particularly CatBoost, surpass traditional tools for ICU mortality prediction in cancer patients.
• Short-term outcomes are primarily driven by acute physiological derangements, not specific cancer characteristics.
• ML-based risk stratification can serve as an objective admission triage tool; external validation is recommended.