Development and External Validation of an Explainable Machine Learning Model to Identify Positive Dysphagia Screening Results in People with Dementia: A Multicenter Cross-Sectional Study.

OBJECTIVE

To develop and externally validate an interpretable machine learning model for identifying the likelihood of a positive Standardized Swallowing Assessment (SSA) result in people with dementia, particularly in settings where instrumental swallowing assessments are not readily available.

METHODS

Candidate variables included demographic, physical, oral-health, and mental health variables routinely available in clinical practice. Least absolute shrinkage and selection operator (LASSO) regression was used for variable selection. Four machine learning models were developed and compared: logistic regression, support vector machines, random forest, and AdaBoost. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), sensitivity, calibration, and decision curve analysis. SHAP analysis was applied to enhance model interpretability. A web-based likelihood assessment tool was developed to estimate the individual probability of a concurrent positive SSA result.

RESULTS

All four models showed good discrimination. AdaBoost achieved an AUC of 0.918 in internal cross-validation and AUCs of 0.900 and 0.889 in the domain and temporal external validation cohorts, respectively, with sensitivities ranging from 0.879 to 0.901. Since sensitivity was prioritized for case finding, AdaBoost was selected for model interpretation and online-tool development. SHAP analysis ranked body mass index as the leading contributor to the AdaBoost output, followed by number of teeth, eating ability, dietary type, and Clinical Dementia Rating score.

CONCLUSION

The model showed good performance for identifying people likely to have a concurrent positive SSA screening result. The web-based tool may support case finding and referral for further swallowing assessment, but it should not be used as a standalone diagnostic or prognostic instrument. Prospective and geographically diverse validation is required before routine clinical implementation.