Machine Learning-Based Prediction of Planned Home Care Follow-Up Intervals in Older Adults: Evidence From the YASAM Observational Cohort.

OBJECTIVES

To develop and validate machine learning models for predicting clinician-determined follow-up interval categories in home care services among older adults enrolled in the YASAM (Healthy Aging Team Supported Home Care Services) program.

DESIGN

Secondary analysis of a multicenter observational cohort study.

SETTING AND

PARTICIPANTS

A total of 4783 community-dwelling adults aged ≥80 years enrolled in the YASAM program between 2023 and 2024.

METHODS

The primary outcome was follow-up frequency time, defined as the planned interval between consecutive face-to-face home care visits determined by an interdisciplinary geriatric care team following comprehensive geriatric assessment and used as the reference standard for model development. Follow-up frequency time was categorized into short (≤30 days), medium (31-90 days), and long (>90 days) follow-up intervals. Seventeen candidate predictors representing demographic, clinical, cognitive, functional, nutritional, mood-related, frailty, and physical performance domains were evaluated. After correlation-based feature selection, 8 core predictors were retained. Missing data were handled using feature-wise imputation. Random forest, extreme gradient boosting, and feed-forward artificial neural network models were developed using a 70/15/15 training-validation-test split. Model performance was assessed using accuracy, precision, recall, F1 score, and confusion matrices.

RESULTS

Extreme gradient boosting demonstrated the highest classification performance, achieving an accuracy of 83%, a precision of 0.83, a recall of 0.82, and an F1 score of 0.82. Random forest achieved an accuracy of 81%, whereas the neural network achieved an accuracy of 78%. Misclassifications occurred predominantly between adjacent follow-up interval categories. SHapley Additive exPlanation-based analyses identified physical performance, cognitive function, nutritional status, and frailty-related measures as major contributors to follow-up interval prediction.

CONCLUSIONS AND IMPLICATIONS: Machine learning models demonstrated substantial predictive performance with interdisciplinary team-determined follow-up interval categories among community-dwelling older adults. Explainable artificial intelligence analyses highlighted the contribution of multidimensional geriatric assessment domains to individualized follow-up planning. These findings support the potential role of artificial intelligence-assisted decision support in home-care scheduling, although external validation is required before routine clinical implementation.