Interpretable machine learning model for predicting the risk of dysphagia in patients after cardiac surgery

Dysphagia is recognized as one of the most common severe complications following cardiac surgery, with the potential to result in adverse clinical outcomes, including aspiration pneumonia, malnutrition, and prolonged hospitalization. Early identification of patients at high risk for postoperative dysphagia is critical for implementing targeted preventive interventions and improving clinical outcomes. However, traditional risk assessment approaches have inherent limitations in accurately predicting this complication. Accordingly, this study was designed to develop and validate a machine learning-based risk prediction model for the accurate identification of patients at high risk for developing dysphagia following cardiac surgery.

A prospective observational cohort study design was used to recruit 573 patients undergoing cardiac surgery at Guangdong Provincial People's Hospital between September 2024 and July 2025. Predictor variables included patient demographic characteristics, preoperative risk factors, intraoperative procedural variables, and clinical postoperative indicators. Eight distinct machine learning models were developed for comparison: logistic regression (LR), adaptive boosting (AdaBoost), categorical boosting (CatBoost), extreme gradient boosting (XGBoost), support vector machines (SVM), naive Bayes (NB), k-nearest neighbors (KNN), and gradient boosting machine (GBM). The complete dataset was randomly divided into a 7:3 training-to-test split for model development and validation, respectively. Model performance was evaluated across multiple metrics, including area under the receiver operating characteristic curve (AUC), accuracy, precision, recall, and F1 score. Model calibration performance was assessed using calibration curves, while clinical utility was evaluated through decision curve analysis (DCA). The optimal-performing model was then interpreted and visualized using SHapley Additive exPlanations (SHAP) to enhance model interpretability for clinical implementation.

The incidence of postoperative dysphagia among the study cohort was 23.03%. Among all evaluated machine learning models, the GBM demonstrated superior overall performance, achieving a test set AUC of 0.851 (95% CI: 0.781-0.908). This performance was statistically significantly better than that observed with all other models evaluated in the study. The key predictive factors identified by the model included patient age, diabetes, chronic lung disease, chronic kidney disease, atrial fibrillation, tracheal intubation time, gastric intubation time, and sedative drug use duration. Additionally, the GBM model demonstrated excellent calibration performance and high clinical utility, indicating significant potential for integration into routine clinical practice to enable accurate, individualized patient risk assessment.

The machine learning model developed in this study demonstrated the ability to accurately identify patients at high risk for postoperative dysphagia, providing a robust foundation for implementing targeted preventive interventions in clinical practice. Integration of this validated machine learning model into standard clinical workflows holds significant promise for reducing the incidence of dysphagia-related complications and improving overall postoperative patient outcomes.