JMIR Publications

ABSTRACT

Background:

As the demand for innovative senior care grows alongside a shortage of care workers, personalization is key to optimizing services and enhancing long-term sustainability. This study proposes an adaptive reinforcement learning-based framework to promote precision digital care by dynamically assigning care programs based on individuals’ unique characteristics and evolving needs. Its effectiveness was evaluated through simulation-based experiments comparing multiple allocation methods within an AI-powered care call service for older adults.

Objective:

To develop and evaluate a reinforcement learning-based model for personalizing digital care program allocation to optimize care engagement and health outcomes among low-income older adults living alone.

Methods:

We developed the framework using contextual bandits, specifically Thompson Sampling, to maximize user outcomes. Four program allocation strategies were tested using a synthetic dataset of user features and program attributes. Simulations were conducted over multiple iterations to evaluate how the model adapts over time and optimizes program assignments compared to static methods.

Results:

Four program allocation methods were compared across 100 simulation runs (n=3,000 assignments per run) using two datasets (AI Call: n=1,196; Community Health Survey (CHS): n=72,812): (i) systematic allocation (baseline), (ii) single best program based on population average, (iii) idealized personalized delivery (theoretical upper bound), and (iv) precision digital care using Thompson Sampling. Precision digital care outperformed baseline and population-average approaches, achieving outcomes comparable to the theoretical upper bound. Compared to systematic allocation, call success rates increased by 84.2% (AI Call) and 54.4% (CHS), PHQ-2 depression scores decreased by 32.1% (AI Call) and 41.4% (CHS), and self-reported health scores improved by 19.0% (AI Call) and 22.0% (CHS). It also showed improved learning efficiency, refining assignments dynamically as it learned from user responses.

Conclusions:

Our findings emphasize the importance of personalization in digital care. We plan to refine and validate the model through a publicly funded AI care program for community-dwelling, low-income older adults living alone in Korea. Reinforcement learning offers a scalable and effective approach to advance precision digital care delivery and support future innovations in aging services.