Machine Learning Reliably Predicts Early Dysphagia After Stroke, suggests study
According to a new study, machine learning can reliably identify patients at high risk of early dysphagia following acute ischemic stroke using routinely available clinical variables. Early dysphagia, affecting almost one-third of the patients with acute ischemic stroke, is associated with aspiration, malnutrition, pneumonia, and increased mortality. This study was published in the International Journal of General Medicine by Ye Li. and colleagues.
Despite its significant consequence on outcomes, early detection of dysphagia in acute ischemic stroke patients remains problematic. Conventional bedside screening may fail to accurately identify high-risk patients, especially within the hyperacute phase. Machine learning approaches are an opportunity to incorporate multidimensional data for enhanced prediction, with maintained clinical interpretability.
This cross-sectional study initially recruited 1,041 patients with acute ischemic stroke from two tertiary hospitals. Participants were divided into a nondysphagia group (n = 736) and a dysphagia group (n = 305). The observed early dysphagia incidence was 29.3%. The dataset was randomly divided into a training cohort (n = 728) and an independent testing cohort (n = 313) using a 7:3 split for model development and validation.
Feature selection was carried out by the Boruta algorithm in combination with logistic regression to identify the most relevant predictors. A total of six machine learning models were trained with 10-fold cross-validation. The performance of models was evaluated by AUC-ROC, sensitivity, specificity, accuracy, positive predictive value (PPV), negative predictive value (NPV), F1-score, and Youden's index. Model explainability was obtained by SHapley Additive exPlanations (SHAP) analysis.
Key Findings
The random forest (RF) model thus gave the best overall predictive performance among all models.
The RF model exhibited an AUC-ROC of 0.952, 95% CI: 0.927-0.976 for discrimination between the early dysphagia and nondysphagia patients in the test dataset.
The six most influential predictors contributing to dysphagia risk, according to SHAP analysis, were ADL grade, NIHSS score, multifocal brain lesions, hypoalbuminemia, coronary heart disease, and lesion hemisphere.
Machine learning models, in particular the Random Forest approach, showed excellent performance in identifying early dysphagia in patients with acute ischemic stroke. The risk factors identified and the explainability framework support early warning systems and personalized management strategies, thus offering a promising path to improve post-stroke care and patient outcomes.
Reference:
Li, Y., Yu, S., Yu, X., Tian, B., Tang, J., Qu, H., & Zhang, Y. (2025). Predicting Early Dysphagia in Acute Ischemic Stroke Using an Explainable Machine Learning Model. International Journal of General Medicine, 18, 7341–7356. https://doi.org/10.2147/IJGM.S567157
