Interpretable machine learning-based prediction of mortality in critical cancer patients with delirium: A retrospective cohort study.

Objective: Delirium in cancer patients presents a significant clinical challenge, often leading to increased mortality, prolonged hospital stays, and higher healthcare costs. This study aimed to develop an interpretable and generalizable machine learning (ML) model for early prediction of mortality risk in cancer patients with delirium. Methods: A retrospective cohort study design was employed, utilizing data from the Medical Information Mart for Intensive Care IV (MIMIC-IV) database. Five ML models were subsequently constructed and evaluated. Results: A total of 1893 cancer patients with delirium were included in the analysis, of whom 685 (36.2%) died within 28 days who were admitted to the intensive care unit at Beth Israel Deaconess Medical Center between 2008 and 2022. The Category Boosting (CatBoost) algorithm outperformed other ML models, achieving the highest area under the curve (AUC) on both training and validation datasets. Its robustness was supported by a bias-corrected performance curve closely aligned with the ideal line and the greatest net benefit in decision curve analysis across all threshold probabilities (0-1). The top five predictors of 28-day mortality were high Glasgow Coma Scale and Acute Physiology and Chronic Health Evaluation II scores, use of antibiotics, propofol, and vasopressors. Conclusions: This study developed an optimal and explainable ML model for predicting 28-day mortality in cancer patients with delirium. The CatBoost algorithm demonstrated stable and robust performance, and interpretability analysis highlighted key predictors. These findings may aid early clinical decision-making and targeted interventions for this high-risk population.