Machine learning–based screening tool for predicting the risk of oropharyngeal dysphagia in patients with ischemic stroke

Abstract

Dysphagia frequently complicates recovery after stroke, necessitating effective predictive tools. This study aimed to develop supervised machine learning models to predict dysphagia risk in ischemic stroke patients. We retrospectively analyzed data from 103 ischemic stroke patients, aged over 18 years, using fiberoptic endoscopic evaluation of swallowing (FEES) as the diagnostic standard for dysphagia. Clinical variables included age, sex, EAT-10 score, Functional Oral Intake Scale (FOIS®), and Glasgow Coma Scale. Eleven ML algorithms were evaluated: Logistic Regression, SVM, Gradient Boosting, Random Forest, Decision Tree, K-Nearest Neighbors, AdaBoost, Multilayer Perceptron, XGBoost, LightGBM, and CatBoost. Data was split into 70% training and 30% testing sets with stratified sampling. Hyperparameter optimization was via grid search and 5-fold stratified cross-validation. Model performance was measured by AUC, accuracy, precision, recall, and F1-score; 95% confidence intervals were estimated by bootstrapping. Predictive models effectively distinguished dysphagia cases. Gradient Boosting and CatBoost achieved the highest AUC values (AUC = 0.99) on the test set. Logistic Regression performed consistently across test and cross-validation sets, achieving an AUC of 0.95 [95% CI: 0.86–1.00] and precision of 0.91 [95% CI: 0.80–1.00] in the test set, and high values in cross-validation (AUC = 0.93 [95% CI: 0.84–1.00]; precision = 0.92 [95% CI: 0.81–0.99]). Key predictors were FOIS®, EAT-10, and age. Machine Learning models showed promise for dysphagia screening post-stroke, with Gradient Boosting, CatBoost, and Logistic Regression showing strong clinical potential for decision support and early referral.