Development and validation of an interpretable machine learning model for predicting enteral nutrition-associated diarrhea in ICU patients: a multicenter study

Abstract

Background ICU patients receiving EN are at high risk of ENAD, which may adversely affect clinical outcomes and increase mortality risk. Early identification of patients at high risk of ENAD may facilitate timely risk assessment and the implementation of individualized nutritional management. This study aimed to develop and externally validate an interpretable ML-based model for predicting ENAD risk in critically ill patients. Methods This retrospective multicenter study included critically ill patients receiving EN at the First Affiliated Hospital of Jinzhou Medical University between January 2024 and December 2025. An independent external validation cohort was obtained from Shangrao People’s Hospital during the same period. Twelve ML models were developed using R software. Model performance was evaluated using AUC, accuracy, precision, NPV, recall, and F1 score. Calibration curves and DCA were used to assess calibration and clinical utility. SHAP analysis was performed to visualize feature importance. Results Among the 12 models, RF showed the best discriminative performance, with an AUC of 0.811 (95% CI: 0.769–0.853) in the test set and 0.808 (95% CI: 0.759–0.856) in the external validation cohort. Conclusion The externally validated random forest model showed acceptable discriminative performance and provided an interpretable assessment of the contributions of individual features. The model may support static early risk stratification of ENAD among critically ill patients receiving enteral nutrition. However, the available evidence remains insufficient to support its routine clinical implementation. Further validation in prospective studies, independent populations, and longitudinal clinical datasets is required.