Development and validation of an explainable machine learning model for predicting prognosis in sepsis patients with a history of cancer who were admitted to the intensive care unit

BackgroundSepsis is the leading cause of mortality in critically ill cancer patients; however, traditional prognostic models fail to capture the complexity of their immune and physiological interactions.MethodsThis retrospective study analyzed electronic health records from the Medical Information Mart for Intensive Care IV database, including the records of patients with sepsis who had a documented history of cancer and were admitted to the intensive care unit. A two-step feature selection approach, combining least absolute shrinkage and selection operator regression and recursive feature elimination, was used to identify key prognostic variables. Eight machine learning algorithms, such as random forest and extreme gradient boosting, were trained and evaluated using five-fold cross-validation. Model performance was assessed using the area under the receiver operating characteristic curve value, Brier scores, sensitivity, and specificity. SHapley Additive exPlanations, Partial Dependence Plots, and break down algorithms were applied to enhance model interpretability.ResultsThe final cohort included 3364 patients admitted to the intensive care unit. Nonsurvivors had significantly higher illness severity scores (Acute Physiology Score III and Sequential Organ Failure Assessment) than survivors (p < 0.001). Among the tested models, the random forest model demonstrated superior performance, achieving the highest area under the receiver operating characteristic curve value (0.78; 95% confidence interval: 0.76-0.80) and the lowest Brier score (0.15), indicating strong predictive accuracy.ConclusionsThis study developed machine learning models for predicting in-hospital mortality in sepsis patients with a history of cancer, leveraging the Medical Information Mart for Intensive Care IV database for comprehensive risk assessment.