JMIR Publications

ABSTRACT

Background:

Wearable devices support self-monitoring but rarely deliver the adaptive, individualized coaching that translates monitoring into sustained behavior change. Large language models (LLMs) could supply that coaching layer at low marginal cost, yet randomized evidence testing LLM-enhanced coaching against an active digital comparator, with a free-living behavioral outcome, is scarce.

Objective:

This study aimed to determine whether adding an LLM-enhanced conversational coach (HabitBot) to a blood pressure smartwatch increases free-living physical activity more than the same smartwatch combined with a standard companion app and a nurse-written exercise prescription in physically inactive adults with prehypertension, and to examine secondary physical activity, anthropometric, engagement, acceptability and theory-aligned behavioral outcomes.

Methods:

We conducted a 12-week, individual-level, 1:1, assessor-blind, parallel-group exploratory randomized controlled trial in Beijing, China. A total of 46 physically inactive adults aged 18-65 years with high-normal blood pressure were randomized to HabitBot or to an active comparator. HabitBot combined GPT-4o conversational coaching delivered through WeChat, progressive action and coping planning, context-linked cues, a blood pressure smartwatch and an optional peer-support group. The comparator received the same smartwatch, a standard wearable companion app and 1 nurse-written individualized exercise prescription. The primary outcome was mean daily smartphone-recorded steps at 12 weeks, analyzed by analysis of covariance adjusted for baseline. Secondary outcomes were questionnaire-assessed physical activity; body composition, blood pressure, theory-aligned behavioral determinants, engagement, acceptability and harms were exploratory.

Results:

All 46 randomized participants completed the 12-week assessment and were analyzed as assigned. HabitBot increased mean daily steps by 1397.6 steps/day relative to the comparator (95% CI 399.4-2395.8; P=.007). The estimate was unchanged across heteroscedasticity-robust, covariate-adjusted, change-score, randomization-inference, bootstrap and leave-one-out sensitivity analyses. Vigorous, moderate and total physical activity favored HabitBot; blood pressure did not differ. Weekly reach was sustained at 21-23 of 23 participants, although interaction volume declined. Behavioral determinants favored HabitBot after false-discovery-rate correction, but these self-reported outcomes were materially constrained by directional baseline imbalance and unblinded self-report. No serious adverse events occurred.

Conclusions:

LLM-enhanced coaching added to a wearable increased short-term physical activity beyond wearable access and standard app support. The trial was small, the intervention multicomponent, and the behavioral mechanism unresolved; adequately powered, component-separating trials are needed before implementation. Clinical Trial: Chinese Clinical Trial Registry ChiCTR2400085073; https://www.chictr.org.cn/showprojEN.html?proj=215626