An evaluation based on explainable machine learning: analysis of risk factors for enteral nutrition-related diarrhea in patients with severe stroke
TL;DR
The results suggest that, in addition to focusing on the patients' clinical condition and nutritional interventions themselves, clinicians should also pay attention to the regulation of ward temperature and the proper management of relevant medications.
Abstract
Objective This study systematically examines the clinical factors associated with enteral nutrition-related diarrhea (ENAD) in patients with severe stroke and uses interpretable machine learning methods to assess the strength of these associations. Methods This retrospective study included 766 patients with severe stroke from two hospitals between January 2021 and December 2025; data were sourced from electronic medical records and nursing notes. The study compared eight machine learning algorithms: logistic regression (Logistic), support vector machines (SVM), random forests (RF), extreme gradient boosting (XGBoost), lightweight gradient boosting (LightGBM), classification gradient boosting (CatBoost), k-nearest neighbors (KNN), and decision trees (DT); model performance was evaluated using 10-fold cross-validation (CV); The optimal model was selected, with SHAP (Shapley Additive exPlanations) analysis used to provide feature importance and the Delong test for feature selection; the final model was evaluated and validated using metrics from calibration curves, decision curve analysis curves, and the model confusion matrix; SHAP analysis was used to interpret the final model. Results A total of 292 patients (38.1%) in the study cohort developed diarrhea. The random forest model was identified as the optimal model and served as the feature selection model. Following the Delong test, 10 variables most closely associated with ENAD were ultimately included: duration of mechanical ventilation, number of days on enteral nutrition (EN) and formula type, C-reactive protein, white blood cell count, hemoglobin, blood urea nitrogen, room temperature, prokinetic agents, and oral potassium supplements. The model performed as follows in terms of retrospective discriminatory power: The AUC was 0.873 (95% CI: 0.826–0.920), with an accuracy of 0.836 (0.790–0.882) and a precision of 0.787 (0.697–0.875). A recall of 0.720 (0.622–0.819), an F1 score of 0.752 (0.672–0.825), and a negative predictive value of 0.859 (0.805–0.912). SHAP analysis revealed a non-linear relationship between the aforementioned features and the risk of ENAD. It should be noted that because some variables (such as duration of mechanical ventilation and number of days on enteral nutrition) reflect cumulative exposure over the entire ICU stay, reported metrics such as AUC may be overestimated and should not be directly interpreted as indicative of predictive performance. The SHAP analysis revealed patterns of association between these characteristics and the risk of ENAD. Conclusion Based on interpretable machine learning methods, this study identified 10 factors—including duration of mechanical ventilation, number of days on enteral nutrition, and C-reactive protein levels—that were most strongly associated with ENAD in the model output in patients with severe stroke. The results suggest that, in addition to focusing on the patients' clinical condition and nutritional interventions themselves, clinicians should also pay attention to the regulation of ward temperature and the proper management of relevant medications. The clinical predictive value of these associations requires further validation through prospective, time-dependent studies.