Clinical Phenotypes, Machine‐Learning Risk Stratification and Treatment Outcomes of Constipation in Stroke Patients Receiving Enteral Nutrition: A Retrospective Study

BACKGROUND: Stroke-related complications and frequent feeding interruptions often compromise nutritional intake, making enteral nutrition (EN) a cornerstone therapy to improve nutritional status and support recovery. Nevertheless, constipation is highly prevalent and clinically consequential among stroke patients receiving EN, particularly in critically ill populations, where enteral nutrition-associated constipation can impede rehabilitation, increase complications and prolong intensive care unit (ICU) stay. AIM: This study aimed to explore potentially important determinants of enteral nutrition-associated constipation (ENAC) in patients with stroke using artificial intelligence approaches, characterise relevant clinical subgroups, compare internally validated machine-learning models and examine how the temporal availability of candidate predictors may influence model performance. STUDY DESIGN: Clinical data from patients admitted to a neurological intensive care unit between December 2019 and December 2023 were retrospectively analysed. K-means clustering was used to identify clinical subgroups. Six algorithms-logistic regression, random forest, gradient boosting tree, XGBoost, LightGBM and Table PFN-3-were compared under temporally restricted and expanded retrospective feature frameworks using five repeats of nested five-fold cross-validation. Performance was assessed using ROC-AUC, average precision, Brier score, calibration and decision-curve analysis. Treatment outcomes were summarised descriptively. RESULTS: Among 362 patients, 233 (64.4%) had documented constipation. Table PFN-3 achieved the highest discrimination in both frameworks. The temporally restricted model yielded a ROC-AUC of 0.629 (95% CI 0.567-0.689), whereas the expanded retrospective model achieved a ROC-AUC of 0.771 (95% CI 0.715-0.820), average precision of 0.857, Brier score of 0.177 and calibration slope of 0.965. The ROC-AUC increased by 0.142 (95% CI 0.090-0.201) after full-hospitalisation information was included. K-means clustering retained three subgroups (K = 3; silhouette coefficient = 0.108), with constipation frequencies of 67.35%, 56.98% and 88.89% in Clusters 0, 1 and 2, respectively. Among evaluable patients, observed relief was 80.3% with glycerin enema, 60.0% with lactulose, 72.7% with combination treatment and 65.7% with neither treatment; these treatment comparisons were descriptive. CONCLUSIONS: Table PFN-3 achieved the strongest internally validated performance, with the expanded retrospective framework combining good discrimination, strong precision-recall performance and close calibration. The 0.142 increase in ROC-AUC after completed hospital-course variables were incorporated quantifies the substantial discriminative information that emerges as hospitalisation evolves. Together with the distinct clinical phenotypes and descriptive treatment findings, these results establish a clinically relevant framework for trajectory-aware bowel monitoring and provide a clear rationale for developing dynamic, time-updated prediction models. RELEVANCE TO CLINICAL PRACTICE: By integrating machine-learning benchmarking, clinical phenotyping and observed treatment outcomes, this study identifies prolonged, treatment-intensive hospital trajectories as a particularly high-burden constipation phenotype and demonstrates that classification performance changes materially as hospital-course information accumulates. These findings support structured reassessment of bowel risk during neurological intensive care rather than reliance on a single static evaluation and provide a practical rationale for future dynamic monitoring tools.