Machine learning in geriatric care: a scoping review of models using multidimensional assessment data

Geriatric assessments capture multidimensional data on physical, cognitive, psychological, and social health, offering opportunities to apply machine learning (ML) to support clinical decision-making in aged care. However, the application of ML to such data has not been systematically synthesised.

To describe the types of data used, purposes, and performance of ML models applied to multidimensional geriatric assessment data.

A scoping review was conducted following the Arksey and O'Malley framework and reported using PRISMA-ScR guidelines. Six databases were searched for peer-reviewed studies (2012--2024) applying ML to multidimensional geriatric assessment data. Data were charted on study design, setting, algorithms, outcomes, and model performance. Methodological quality was assessed using the PROBAST + AI tool.

Forty studies met inclusion criteria, most from high-income countries and community or hospital settings. Frequently modelled outcomes included falls, functional decline, frailty, mortality, and hospital readmission. The most common data domains were comorbidities, cognition, mood, and ADL/IADL. Most studies (25/40) compared multiple algorithms, with the XGBoost emerging as the top-performing model in eight studies. Although internal validation was common, no study performed external validation. Performance metrics varied widely, and methodological limitations such as small sample sizes, lack of calibration, and data imbalance were common.

ML models using multidimensional geriatric assessment data show promise for predicting health outcomes in older adults. However, methodological and reporting limitations hinder clinical translation. Future research should focus on external validation, interpretability, and integration into clinical workflows to ensure models are robust, ethical, and applicable in real-world aged care settings.