Development of machine learning-based models for predicting sarcopenia risk in stroke patients and analysis of associated factors

This study aimed to develop and validate machine learning models to predict sarcopenia risk in stroke patients and to identify the most relevant clinical features associated with its occurrence. In this prospective study, 425 stroke patients were enrolled between October 2024 and April 2025. Patients from Kunming First People's Hospital (n = 308) formed the training cohort, while those from Kunming Yan'an Hospital (n = 117) comprised the validation cohort. Feature selection was performed using a random forest algorithm. Five machine learning models-logistic regression, decision tree, random forest, naïve Bayes, and gradient boosting-were developed and evaluated using accuracy, recall, precision, specificity, F1 score, and area under the curve (AUC). SHapley Additive exPlanations (SHAP) were applied to the best-performing model to interpret key predictors. Of the 425 patients, 145 (34.1%) were diagnosed with sarcopenia. Significant differences (P < 0.05) were observed between the sarcopenia and non-sarcopenia groups across multiple clinical variables. The random forest model demonstrated the best predictive performance across all metrics. SHAP analysis revealed BMI, serum albumin, age, uric acid, creatinine, hemoglobin, calcium, triglycerides, C-reactive protein, total protein, and urea as the most influential predictors, reflecting nutritional, metabolic, and inflammatory status. The random forest model achieved superior performance in predicting sarcopenia risk after stroke. Combining machine learning with SHAP interpretability offers a robust and explainable framework for early identification and personalized management of post-stroke sarcopenia, supporting precision geriatric care.