A Hybrid Rule- and Large Language Model-Based Embodied Voice Assistant (GRACE) for Cognitive Stimulation in Older Adults: Usability Study Assessing Technical Feasibility, Technology Acceptance, and Working Alliance.

BACKGROUND: The health and economic burden of dementia has led the World Health Organization to recognize it as a public health priority. Although there currently does not exist a cure for dementia, there are multiple interventions aimed at preventing the risk of dementia and improving the quality of life of people with dementia. Voice assistants (VAs), particularly those using large language models (LLMs), have emerged as promising tools to deliver these interventions to older adults due to their accessible and natural interface. OBJECTIVE: This pilot study aimed to evaluate the technical feasibility (ie, functional performance and usability) and user acceptance of the embodied rule-based and LLM VA GRACE, as well as the perceived strength of the collaborative relationship or working alliance, between GRACE and healthy older adults during the delivery of cognitive stimulation interventions. METHODS: A pilot study was conducted with 21 healthy German-speaking adults aged 60 years and older. Participants interacted with GRACE in a laboratory setting for 10-15 minutes. The interaction involved a structured cognitive stimulation session using rule-based and LLM components. Data were collected using pre- and postinteraction questionnaires and semistructured interviews. Quantitative analysis included descriptive statistics and Wilcoxon signed rank tests. Qualitative data were analyzed thematically. RESULTS: Participants rated GRACE positively, with statistically significant scores above neutral (P<.001 for perceived ease of use, usefulness, enjoyment, and working alliance; P=.009 for perceived control; and P=.009 for intention to continue interacting). Thematic analysis revealed that GRACE was perceived as easy to understand and unambiguous, friendly, and supportive, with intervention components viewed as enjoyable and appropriately challenging. Areas for improvement included personalization, response delays, and voice quality. CONCLUSIONS: The results suggest that embodied rule-based and LLM VAs like GRACE are feasible and well-received tools for delivering cognitive interventions to older adults. Future iterations will incorporate feedback and extend testing to individuals at risk for dementia.