Service combination patterns and predictors in long-term care: A machine learning approach

Abstract Background Long-term care (LTC) users often receive multiple services in combinations, yet systematic insights into the drivers of different patterns remain limited. This study applies machine learning to identify distinct service combination patterns among LTC users and determine key influencing factors, providing evidence to inform user-centred policy development. Methods We analysed service records from 41,323 LTC users in a Taiwanese county, covering eight categories: home care, day care, adult foster care, community care, professional care, transportation, respite care, and meals on wheels. Methods included: (1) calculating service utilisation frequency; (2) applying K-means clustering to group users by service combination patterns; (3) using Bootstrap LASSO to enhance variable selection robustness; and (4) building logistic regression models to estimate predictor effects. Results Eight distinct service combination types were identified. Thirty-six significant predictors were classified into six domains: Demographic Characteristics (6 predictors), Caregiver Information (7), Health Status Indicators (7), Functional Status and Mobility (9), Living Environment and Social Support (5), and Access and Entry Characteristics (2). These factors contributed to distinguishing different service combination patterns. Conclusions This study offers a novel, data-driven perspective on LTC utilisation by identifying user groups defined by preferred service combinations. The findings support the development of differentiated, needs-based service models, advancing user-centred LTC systems and informing public health policy. Key messages • Applying machine learning, we identified diverse service combination patterns among LTC users, informing targeted service planning and resource allocation. • Predictors of service combination patterns can guide the development of user-centred and needs-based long-term care service models.