Machine Learning Classification of Frailty Using Wearable-Derived Sleep Metrics in Community-Dwelling Older Adults

Frailty is a multifactorial geriatric syndrome, and sleep disturbances have emerged as a potential contributing factor. However, conventional statistical approaches may not adequately capture the complex and nonlinear patterns inherent in wearable-derived sleep data. This study evaluated the feasibility of machine learning (ML) models for classifying frailty using objectively measured sleep features in older adults under free-living conditions.

This cross-sectional study included 90 community-dwelling adults aged ≥65 years. Participants were classified as robust or pre-frail/frail using the Fried frailty phenotype. Sleep data were collected via a wrist-worn device, and sleep stage, duration, and interaction features were extracted. Five ML models (random forest, gradient boosting, categorical boosting, extreme gradient boosting, and logistic regression) were developed, and model performance on an independent test dataset was evaluated using accuracy, F1 score, receiver operating characteristic area under the curve (ROC-AUC), precision-recall area under the curve (PR-AUC), sensitivity, specificity, and Matthews correlation coefficient. Shapley additive explanation (SHAP) analysis was applied to enhance model interpretability.

Among the evaluated models, random forest showed the best overall performance, achieving the highest ROC-AUC of 0.892 and PR-AUC of 0.894. SHAP analysis identified age, cognitive function, rapid eye movement sleep, sleep duration, and interaction features reflecting sleep characteristics as important contributors to frailty classification.

Wearable-derived sleep monitoring combined with interpretable ML may provide a feasible and noninvasive approach for early frailty risk identification in older adults. Sleep features, together with age and cognitive function, may serve as meaningful indicators of frailty risk in older adults.