An explainable multimodal ultrasound machine-learning model for identifying clinically defined oropharyngeal dysphagia: prospective two-center development and validation

Oropharyngeal dysphagia is increasingly prevalent in aging populations and in patients with neurological disorders, and is associated with aspiration, malnutrition, and pneumonia. Videofluoroscopic swallowing study (VFSS) and fiberoptic endoscopic evaluation of swallowing (FEES) remain reference-standard examinations, but their use may be limited in settings requiring repeated or point-of-care assessment. Portable multimodal ultrasound may provide complementary biomechanical information, although standardized ultrasound-based assessment protocols remain under development. In this prospective two-center study, we enrolled adults with clinically defined dysphagia and healthy controls. Dysphagia status was determined using a standardized clinical framework incorporating swallowing-related assessment scales and bedside swallowing evaluation. All participants underwent a standardized multimodal ultrasound protocol for biomechanical swallowing assessment. Feature selection was performed using a multistage procedure, and seven machine-learning classifiers were developed and compared. Seven classifiers were trained in the training cohort and compared in the internal validation cohort; the locked final model was subsequently evaluated in the independent external test cohort (n = 28). Performance was measured by discrimination, calibration, and clinical utility, with model interpretability analyzed using SHapley Additive exPlanations (SHAP). Among 236 enrolled participants, 114 had clinically defined dysphagia, mainly related to amyotrophic lateral sclerosis and stroke. Six biomechanical and clinical features were selected for model development, with age included as a covariate. The support vector machine (SVM) model showed the strongest overall performance, with an area under the receiver operating characteristic curve (AUC) of 0.993 (95% confidence interval, 0.976–1.000) in internal validation and 0.938 in the independent external cohort. SHAP analysis identified prolonged tongue contraction time and reduced movement speeds as the most influential predictors. A machine-learning model integrating quantitative multimodal ultrasound features may support the identification of clinically defined oropharyngeal dysphagia as an adjunct to conventional assessment. This portable ultrasound-based approach may facilitate point-of-care risk stratification, but should not be interpreted as replacing VFSS, FEES, or clinical judgment. Further validation in larger and etiologically diverse cohorts is needed to confirm generalizability and clinical utility.