JMIR Publications

ABSTRACT

Background:

Caregivers of people living with dementia (PLwD) face significant stress, particularly when verifying whether tasks are truly completed, despite the use of digital reminder systems. While PlwD may acknowledge reminders, caregivers often lack a reliable way to confirm task adherence. Generative AI, such as GPT-4, offers a potential solution by automating task verification through follow-up questioning and supporting caregiver decision-making.

Objective:

This feasibility study evaluates an AI-powered task verification system integrated with a digital reminder framework for PLwD. Specifically, it examines (1) the effectiveness of GPT-4 in generating high-quality follow-up questions that help verify whether tasks were actually completed, (2) the accuracy of an AI-driven response flagging mechanism in identifying tasks requiring caregiver intervention, and (3) the role of caregiver feedback in refining system adaptability.

Methods:

A theoretical AI-powered task verification pipeline was designed to enhance digital reminders by generating tailored follow-up questions, analyzing responses, and categorizing concerns. Each follow-up question corresponded to a specific reminder sent through the digital system, aiming to assess whether the task was genuinely completed. To test its feasibility, a simulated AI-powered pipeline was implemented using an anonymized dataset of 64 reminders. GPT-4 generated follow-up questions with and without additional contextual information about PLwD routines. A response classification system flagged task completion as High, Medium, or Low concern, based on response clarity and task urgency. Simulated caregiver feedback was incorporated to refine question quality and improve system adaptability over time.

Results:

Contextual information and caregiver feedback significantly improved the clarity, specificity, and relevance of AI-generated follow-up questions. The response flagging mechanism demonstrated high accuracy, particularly for critical tasks such as safety-related reminders. However, subjective or non-urgent tasks posed classification challenges. Caregiver input iteratively enhanced system performance, ensuring a balance between automation and human oversight.

Conclusions:

This study demonstrates the feasibility of integrating generative AI into dementia care for task verification and caregiver support. AI-generated follow-up questions provide a structured way to confirm whether tasks were truly completed after a digital reminder was acknowledged. The findings suggest that context-aware AI-generated prompts, combined with iterative caregiver feedback, improve task verification accuracy, reduce caregiver stress, and enhance PLwD support. Future research should focus on real-world implementation, longitudinal usability, and scalability to optimize AI-driven dementia care interventions.