Related Experiment Videos
Development and validation of an interpretable machine-learning model for enteral nutrition-associated diarrhea in
1Department of Critical Care Medicine, The First College of Clinical Medical Science, Yichang Central People's Hospital, China Three Gorges University, Yichang, Hubei, China.
Background:
Enteral nutrition-associated diarrhea (ENAD) is a frequent gastrointestinal complication during enteral nutrition (EN) in critically ill patients. In ICU patients with ischemic stroke, neurological impairment, dysphagia, reduced consciousness, immobility, infection, and intensive care interventions may further impair gastrointestinal tolerance; however, disease-specific prediction tools are scarce.
Objective:
To develop and internally validate an interpretable machine-learning model for individualized ENAD prediction in ICU patients with ischemic stroke receiving EN.
Methods:
This single-center retrospective cohort study included adult ICU patients with ischemic stroke who received EN at Yichang Central People's Hospital from January 2024 to December 2025. ENAD was defined as Bristol Stool Form Scale type ≥6 plus either ≥3 bowel movements/day or stool output >500 g/24 h after EN initiation. Candidate predictors covered demographic, neurological, nutritional, laboratory, EN-related, medication, and organ-support variables. Missing values were handled within five-fold cross-validation training folds. LASSO logistic regression was used for feature screening, and six algorithms were compared using five-fold cross-validated out-of-fold predictions. Performance was assessed using discrimination, precision-recall performance, calibration-related metrics, confusion-matrix indices, decision-curve analysis, and SHAP interpretation.
Results:
Among 374 patients, 105 developed ENAD (28.1%). Patients with ENAD had higher NRS-2002, APACHE II, NIHSS, and mRS scores; lower GCS scores; longer ICU stay and mechanical ventilation; higher EN infusion rates and CRP; and lower albumin. The random forest model showed the best overall internal performance, with AUC 0.969, AUPRC 0.921, sensitivity 0.943, specificity 0.888, negative predictive value 0.98, and F1 score 0.85 at a threshold of 0.45. SHAP identified EN infusion rate, CRP, NRS-2002 score, albumin, mRS score, APACHE II score, and NIHSS score as leading contributors.
Conclusion:
Pending external validation, the interpretable random forest model can support an early ENAD risk pathway for ICU patients with ischemic stroke receiving EN. High-risk predictions should trigger structured stool monitoring, feeding-rate and delivered-versus-prescribed EN review, medication review, skin protection, fluid-electrolyte surveillance, and early nutrition-support consultation. Future work should prioritize multicenter validation, calibration updating, parsimonious model comparison, and prospective impact testing before routine clinical deployment.