JMIR Publications

ABSTRACT

Background:

Traditional conversational agents (CAs) for health and well-being interventions often suffer from rigid responses and limited adaptability, particularly for individuals with cognitive impairments. The recent integration of Large Language Models (LLMs) offers enhanced natural language understanding and personalization, presenting promising tools for behavior-driven interventions. However, the role of LLM-powered CAs in supporting physical activity and cognitive training remains underexplored.

Objective:

This scoping review explores the role of LLM-based CAs in supporting individuals' physical activity and cognitive training, providing a comprehensive overview and critical evaluation of their impact.

Methods:

We systematically searched Clarivate Web of Science and Elsevier Scopus using precise query strings. From 354 records, 10 papers focusing on LLM-based CA interventions for physical activity or cognitive training were rigorously screened and extracted by six authors. Data synthesis involved both qualitative and quantitative descriptive analysis.

Results:

Our findings indicate a predominant application of LLM-based CAs in physical activity, often adopting coaching roles, while their use in cognitive training is nascent. The design landscape relies on proprietary LLMs and inconsistently documented prompt engineering. Reported outcomes primarily centered on perceived usefulness and engagement, with a notable lack of robust quantitative efficacy evaluations.

Conclusions:

LLM-based CAs show promise for well-being interventions, especially in physical activity. However, their scientific maturation and widespread adoption are hindered by evaluation rigor and critical reproducibility issues unique to LLMs. The black-box nature and continuous evolution of proprietary LLMs, coupled with informal prompt sharing, pose fundamental barriers to replication. This review underscores an urgent need for greater transparency in reporting LLM versions, standardized prompt engineering, and more rigorous open science methodologies for responsible field advancement.