Development and validation of a screening model for dysphagia in the elderly based on acoustic features

Background: Dysphagia is a prevalent and serious condition among the elderly, yet scalable screening tools are lacking. This study aimed to develop and validate an automated machine learning model based on acoustic features for screening dysphagia risk in the elderly. Methods: Adhering to TRIPOD guidelines, we conducted a study in three stages: variable screening, model construction, and evaluation. Audio data (voice, cough, swallow) were collected from the elderly in nursing homes. A modeling dataset (Beijing area, n = 419) was used to screen key features via LASSO regression. Models were built using Logistic Regression, Random Forest, SVM, and XGBoost, with performance evaluated on an internal test set. The best-performing model was subsequently validated on an external dataset (Shijiazhuang area, n = 216). Results: The XGBoost model demonstrated superior performance, with an area under the curve (AUC) of 0.86 in internal validation and an AUC of 0.71 in external validation, showing good discrimination, calibration, and clinical utility. Conclusion: The acoustic feature-based XGBoost model serves as an effective and automated tool for screening dysphagia risk in the elderly. It has the potential to assist healthcare professionals in identifying high-risk individuals for early intervention, thereby improving clinical outcomes.