Interpretable machine learning for accessible dysphagia screening and staging in older adults

Dysphagia in older adults causes serious complications, and efficient and scalable screenings are needed. This prospective multicenter study developed interpretable machine learning (ML) models for the early identification and staging of dysphagia. Nine ML models were built using the clinical data from 1,235 patients and externally validated on 720 patients. All patients were older adults from seven Suzhou hospitals whose dysphagia was confirmed via videofluoroscopic swallowing studies. Features were selected via random forest, and model interpretability was analyzed with SHapley Additive exPlanations (SHAP). The CatBoost model achieved an area under the receiver operating characteristic curve (AUC) of 0.914 for binary classification, while neural network gave AUC 0.884 for multiclass classification. External validation confirmed robustness (binary AUC, 0.909 and multiclass macro-AUC, 0.860). SHAP identified ten core features-oral/pharyngeal function influenced all stages, and masticatory/phonatory features acted selectively. A web application was created accordingly to facilitate real-time screening and stratify dysphagia patients.