In Korea, researchers are actively developing a range of government-funded digital care programs aimed at improving the well-being of older adults. These programs are administered by local authorities and are currently tailored to specific groups of older adults in distinct regions. As these initiatives progress, the accumulation of knowledge and data is expected to pave the way for large-scale implementation in the near future. This study focused on the critical role of personalization in enhancing the effectiveness of digital care. While personalization has not yet been fully integrated, demographic data on recipients of digital care programs is available. Leveraging this information, we conducted simulation studies that modeled the baseline characteristics of user populations, simulating the potential impact of personalized interventions.

To evaluate the potential of personalized digital care, we conducted a series of simulations mimicking the delivery of care programs to older adult residents in a community setting.

To investigate the potential of personalized program allocation under different scenarios, we constructed synthetic user profiles representative of the 2 populations described above. User profiles were defined by a combination of 7 binary demographic and health-related features: age (older or younger), sex (male or female), region (rural or urban), economic status (unstable or stable), presence of multiple chronic conditions (present or absent), functional difficulty (present or absent), social isolation (high or low), cognition (intact or not intact), and living arrangement (alone or with others). The generation process differed depending on the simulation set:

For simulations based on the AI Call dataset: Synthetic user profiles were generated by sampling directly from the characteristics of about 1000 users in the raw AI Call dataset. This ensured the simulated users closely mirrored the specific population engaging with that service. For simulations based on the CHS dataset: since only aggregated distributional information (eg, percentage breakdowns for age, sex, and region) was available from the CHS summary data, synthetic user profiles were generated differently. Each user characteristic was randomly generated independently according to the proportions reported in the CHS data. This method produced a synthetic population whose overall marginal distributions matched the CHS population structure, assuming independence between characteristics based on the available information. Similarly, we created a pool of programs with varying attributes (content type, frequency, duration, and voice type). Parameters defining individuals (based on the methods above) and the range of programs were set to represent typical variations observed in Korean older adult care contexts.

A critical component defining each simulation scenario is the relationship between a user's characteristics, the assigned program attributes, and the resulting outcome or reward. This relationship models the differential program effectiveness, which personalization aims to exploit. The construction of this relationship also differed between the two simulation sets: (1) For simulations based on the AI Call dataset, the parameters (coefficients) governing the relationship between user and program features and outcomes were informed by observed patterns in the actual AI Call data. Specifically, the direction (sign) of main effect coefficients was set to match observed trends (eg, whether rural users in the data had higher or lower call success). For simplicity, the magnitude of these data-informed coefficients was held constant; and (2) for simulations based on the CHS, since the CHS data lacked information on how user characteristics relate to outcomes from specific digital care programs, the coefficients defining the relationship between user and program features and outcomes were assigned randomly. This simulated outcome variability without imposing specific assumptions based on the limited data.

In both simulation sets, to explicitly model the potential for personalization, several user-program interaction effects were included. These were based on domain knowledge and plausible scenarios. Importantly, these interaction effect scenarios were applied consistently across all 3 outcome measures. Specific examples of modeled interactions include scenarios where:

Within the simulation framework described above, we modeled 3 key outcomes to assess program effectiveness. The generation of these outcomes for each user-program interaction followed the relationships defined in “Setup for the Relationship Between Reward and Context” (differing based on whether it was an AI Call or CHS-based simulation). We compared the average value of each outcome measure over 100 repetitions, each involving 3000 user-item interactions (call selections) with a randomly shuffled sequence of users. The three specific outcome measures we used are:

Four allocation methods were evaluated based on 2 metrics. The first metric was the distribution of the average outcome across simulation runs. For each of the 3 performance measures, call success, PHQ-2 score, and SRH, the average value of the outcome for each of the 4 allocation methods was calculated. This represents the average scores of the subjects who received the care call according to each allocation rule.

The second metric was average cumulative regret over time, or with the increasing number of subjects. Cumulative regret represents the loss incurred from selecting a suboptimal arm over time. For the first 2 allocation methods, regret never decreases, and cumulative regret increases linearly over time. For Method 3, both regret and cumulative regret are zero. For the proposed method, RL-based precision digital care, if the method learns the optimal arm, cumulative regret tapers off after increasing initially. The tapering nature indicates that the method is learning the optimal arm.

This study was reviewed and determined to be exempt from ethical approval by the Institutional Review Board of Seoul National University (E2408/004-007; August 14, 2024), as it used publicly available and deidentified data that did not involve direct interaction with human participants. The study analyzed anonymized AI care call data and aggregated data from the Community Health Survey, and no personally identifiable information was accessed or retained. Because the study used deidentified secondary data and simulation-based analyses, informed consent from participants was not required. All procedures were conducted in accordance with the Declaration of Helsinki.

Our findings highlight the potential of precision digital care with Thompson Sampling in optimizing digital care service allocation for older adults. Through simulation studies, we demonstrated that personalized program allocation leads to improved care outcomes compared to nonpersonalized methods. Specifically, the precision digital care approach outperformed both systematic and single-best allocation approaches, achieving results close to the idealized theoretical upper bound. This underscores the importance of hyper-personalization in digital health care, particularly in aging populations facing growing care demands and workforce shortages.

The outcomes evaluated in this simulation study represent key dimensions of effective care delivery for older adults living alone. Call success indicates sustained engagement with care services and regular monitoring, both critical for timely detection of health or safety concerns and prevention of unattended deaths- an emerging social problem in Korea [20]. PHQ-2 scores represent depressive symptoms that affect daily functioning, social participation, and self-care capacity, while self-reported health reflects subjective well-being and perceived ability to age in place.

In real-world implementation, improvements in these outcomes would support the delivery of responsive, person-centered care. Importantly, precision digital care is designed to augment (not replace) human care relationships. In Korea’s context, direct care workers often face overtime burdens and emotional exhaustion, contributing to workforce shortages [2,3]. The proposed approach optimizes routine care allocation through AI-driven monitoring, allowing human workers to focus on complex cases requiring empathy, clinical judgment, and relational continuity. This human-AI collaboration model addresses workforce sustainability while preserving the irreplaceable value of human connection in care relationships. Rather than automating care relationships, the system automates administrative burden, the repetitive monitoring tasks that overwhelm workers, enabling them to provide higher-quality human care where it matters most. Together, these outcomes support the goal of enabling older adults to age with dignity and autonomy while ensuring appropriate care allocation.

These findings align with previous research emphasizing the importance of personalization in digital health interventions for older adults. A scoping review of digital health interventions targeting older populations found that personalized features improve self-management and psychosocial health, but also noted that experimental studies validating these effects are scarce [21]. Our study contributes to this gap by providing a systematic, AI-driven approach to personalization, demonstrating through simulation that RL can effectively optimize program allocation based on user characteristics. Unlike previous work, which primarily relies on static program designs, our approach adapts dynamically, ensuring that interventions are continually optimized over time.

Our study also relates to research exploring contextual bandit algorithms in mobile health settings, where RL has been used to optimize treatment timing and intervention strategies [22]. Existing studies have demonstrated the feasibility of such approaches in personalized patient engagement and behavior change interventions [23]. However, most prior applications focus on mobile health and preventive care rather than digital older adult care services. Our findings extend these applications by showing that contextual bandits can optimize not just treatment timing but also care program selection for older adults in government-supported digital health initiatives. This represents a novel application of RL in aging-related digital interventions, offering a scalable solution for optimizing care programs at the population level.

Beyond demonstrating the effectiveness of RL for digital care optimization, our findings have broader implications for the scalability of AI-driven decision models in government-funded care programs. As governments invest in digital health solutions to address aging populations and workforce shortages, integrating AI-driven program allocation systems can improve care efficiency without requiring extensive prelabeled datasets [24]. The Korean AI care call program exemplifies this trend, yet current models lack individualized program allocation strategies. By applying the proposed precision digital care approach for personalization of care in such settings, policymakers could transition from standardized digital interventions to truly personalized care, ensuring that digital services maximize impact and cost efficiency. More broadly, the integration of AI-driven decision-making into digital care services aligns with global trends toward precision health and personalized medicine, advocating for a shift away from static care delivery models.

Despite these promising findings, several limitations must be acknowledged. This study relied on synthetic data and simulations, rather than real-world implementation. Although the data was constructed using existing survey datasets and AI care call user characteristics, real-world trials are necessary to validate these findings in practice. Additionally, the current model was developed within the Korean digital care context, and further research is needed to assess the adaptability of precision digital care in different health care systems and cultural settings.

This study’s focus on low-income, socially isolated older adults living alone reflects an intentional equity commitment. These populations face compounded vulnerabilities, limited access to family support, health care services, and social resources, placing them at increased risk of adverse outcomes [25]. Precision digital care has particular promise for reducing geographic and functional disparities. Our simulations incorporated domain-informed interaction effects in which rural users and those with functional limitations were modeled as benefiting from different care configurations. For example, based on observed patterns in Korean older adult care contexts, rural older adults were modeled as preferring more frequent calls (potentially compensating for limited in-person service access), while those with functional difficulties were assumed to benefit from longer interaction times (allowing more comprehensive needs assessment). Similarly, older adults with cognitive limitations were modeled as engaging better with simpler check-in style calls rather than complex health information content.

By identifying and responding to these differential needs, precision digital care can help bridge urban-rural care gaps, accommodate varying cognitive and functional capacities, and ensure that digital care innovations reduce rather than exacerbate existing inequities. However, realizing this potential requires careful attention to digital literacy barriers, technology access, and culturally appropriate implementation, critical considerations for future real-world validation studies.

Beyond technical and operational concerns, AI-driven personalization in digital health raises ethical considerations, such as potential bias in program allocation, equity in digital access, and privacy concerns related to AI-driven recommendations [26]. Future studies should explore fairness-aware RL approaches to ensure equitable care distribution, particularly for older adults who may have limited digital literacy or reduced access to technology. Furthermore, while RL-based precision digital care optimizes individualized program assignments, its integration into broader multilevel care frameworks, such as clinical care coordination, home-based services, and telemedicine ecosystems, remains an open challenge. Ensuring seamless interoperability with existing digital health platforms will be essential for widespread adoption.

To address these limitations, we plan to conduct real-world validation studies in a publicly funded AI care call program in Korea, assessing precision digital care’s impact on user satisfaction, health outcomes, and long-term engagement, with an expanded set of objective measures (eg, health care use and device-measured activity) to complement self-reports. Although equity was not quantitatively evaluated in this study, the intervention setting, a publicly funded AI care call program for low-income, socially vulnerable older adults living alone, reflects an inherent focus on underserved populations. Future research should also explore hybrid AI models, integrating deep learning and RL to further refine personalized digital interventions. Additionally, a critical next step is to examine how RL-based precision digital care performs across diverse populations, ensuring that AI-driven personalization remains both effective and equitable across different health care systems.

This study introduced precision digital care, an AI-driven RL model for optimizing personalized digital care service allocation. Our simulation-based evaluation demonstrated that precision digital care significantly improves care outcomes by dynamically adapting interventions to individual user characteristics. These findings highlight the potential of AI-driven personalization in enhancing the efficiency and effectiveness of digital health interventions, particularly in aging populations where scalable, individualized care is increasingly needed.

As digital care programs expand, ensuring personalized, data-driven decision-making will be critical in maximizing their impact. Precision digital care offers a scalable approach to deliver dignified, person-centered care at scale. By using RL to understand and respond to individual needs, the proposed framework can help older adults maintain autonomy, social connection, and well-being while supporting sustainable care systems. The ultimate goal is not simply algorithmic optimization but improving lived experiences and quality of life in later years through adaptive, humane digital care that honors individual dignity and preferences. However, further real-world validation is required to confirm the effectiveness of precision digital care beyond simulations and to address practical implementation challenges.

Future research should focus on scaling and adapting precision digital care to diverse health care settings, ensuring ethical AI deployment, and integrating hybrid AI models to enhance personalization. Establishing fair, transparent, and explainable AI systems will be key to promoting trust and equitable access in digital health services for aging populations worldwide.