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

Area of Science:

• Gerontology
• Medical Informatics
• Biomedical Engineering

Background:

• Dysphagia in older adults leads to severe health issues, necessitating effective screening methods.
• Current screening approaches may lack efficiency and scalability for the aging population.

Purpose of the Study:

• To develop and validate interpretable machine learning (ML) models for early dysphagia identification and staging in elderly individuals.
• To create a practical tool for real-time dysphagia screening and patient stratification.

Main Methods:

• Prospective multicenter study involving 1,235 patients for model development and 720 for external validation.
• Utilized nine ML models, feature selection via random forest, and SHapley Additive exPlanations (SHAP) for interpretability.
• Validated models using videofluoroscopic swallowing studies (VFSS) confirmed dysphagia diagnoses.

Main Results:

• CatBoost model achieved high accuracy (AUC 0.914) for binary dysphagia classification; neural network performed well (AUC 0.884) for multiclass staging.
• External validation confirmed model robustness (binary AUC 0.909, multiclass macro-AUC 0.860).
• SHAP analysis identified key features, highlighting oral/pharyngeal function, masticatory, and phonatory aspects influencing dysphagia stages.

Conclusions:

• Interpretable ML models offer a robust and scalable solution for early dysphagia detection and staging in older adults.
• The developed web application facilitates clinical implementation for real-time screening and patient management.
• Understanding key predictive features enhances clinical insights into dysphagia pathophysiology.