A Machine-Learning Model Based on Clinical Features for the Prediction of Severe Dysphagia After Ischemic Stroke

This study aimed to construct machine-learning models for prediction of severe dysphagia after ischemic stroke based on clinical features and identify significant clinical predictors.

Patients hospitalized with dysphagia after ischemic stroke in Affiliated Hospital of Jiangnan University were retrospectively analyzed and randomly divided into training and validation sets at a ratio of 7:3. Additional patients from Huai'an Hospital were selected as test set. 19 relevant clinical characteristics were collected. According to the water swallowing test (WST), patients were divided into severe dysphagia group and non-severe dysphagia group. K-nearest neighbor (KNN), decision tree (DT), random forest (RF), support vector machine (SVM), light gradient boosting machine (LGBM), and extreme gradient boosting (XGBoost) were applied to predict severe dysphagia. Receiver operating characteristic (ROC) curves were plotted, the area under the ROC (AUC) was calculated to assess predictive power, and DeLong's test was used to compare the AUCs among six models. Finally, an optimal model was obtained, and significant clinical predictors of severe dysphagia after stroke were screened.

A total of 724 patients were enrolled, 422 in training set, 182 in validation set and 120 in test set, respectively, with no statistically differences in baseline information (P>0.05). In the training set, the AUCs of KNN, DT, RF, SVM and XGBoost were higher than that of LGBM (P<0.05). In the validation and test sets, the AUCs of XGBoost were also higher. The performance metrics of XGBoost were better in terms of accuracy, precision, recall, and F1-score. Therefore, XGBoost was the best model, with good clinical practicality. Furthermore, the top five features based on XGBoost were NIHSS score, BI, BMI, age and time since stroke onset.

Among all clinical feature-based machine-learning models for the prediction of severe dysphagia after ischemic stroke, XGBoost had the best predictive value.