Machine learning-based predictive tools and nomogram for in-hospital mortality in critically ill cancer patients: development and external validation using retrospective cohorts.

Abstract

Background: The incidence of intensive care unit (ICU) admissions and the corresponding mortality rates among cancer patients are both high. However, the existing scoring systems all lack specificity. This research seeks to establish and validate a prediction model for early forecasting of in-hospital mortality in critically ill cancer patients.

Methods: A retrospective analysis was conducted utilizing data from cancer patients obtained from the eICU and MIMIC-IV databases. The least absolute shrinkage and selection operator method was employed to screen predictive factors and construct six machine learning (ML) models. These models were mainly compared in terms of their predictive performance through area under the curve (AUC) and underwent external validation. Nomograms were developed using multivariate logistic regression (LR) analysis findings. The Shapley Additive exPlanations (SHAP) method was employed to explain the variables within the ML models.

Results: Twelve predictive factors were chosen to develop the ML models. Among these models, the LR model and the eXtreme gradient boosting (XGB) model demonstrated the optimal efficacy. In the external validation cohort, their AUC values reached 0.751 [95% confidence interval (CI): 0.735 - 0.768] and 0.737 (95% CI: 0.720 - 0.754), respectively. Moreover, nomograms and SHAP were employed to explain the variables. Additionally, a user-friendly web-based calculator tool was created.

Conclusions: The LR and XGB models were successfully developed to predict in-hospital mortality in critically ill cancer patients. Their robust predictive ability was demonstrated in the external validation cohorts. This model can assist physicians in clinical decision-making and timely intervention.

Clinical trial number: Not applicable.