JMIR Publications

ABSTRACT

Background:

Large language models (LLMs) are increasingly embedded in conversational agents for cardiometabolic care. These systems could support self-management, but their behaviour change content, delivery mechanisms, and implementation transparency are poorly understood.

Objective:

This scoping review mapped behaviour change techniques (BCTs) used in LLM-driven conversational agents for cardiometabolic prevention and management, described how these techniques are delivered (static, rule-based, or generative), and examined how transparently LLM design, personalisation, and safety features are reported, and summarised what has been reported regarding user experience and preliminary behavioural or clinical outcomes.

Methods:

Following the Arksey and O’Malley framework (as refined by Levac et al) and PRISMA-ScR guidance, we searched PubMed, Web of Science, Embase, CINAHL, APA PsycInfo, IEEE Xplore, ACM Digital Library, arXiv, ClinicalTrials.gov, and the WHO International Clinical Trials Registry Platform for records published between January 1, 2020 and November 30, 2025, with the final search completed on March 25, 2026. Eligible studies reported a patient-facing text- or voice-based conversational agent for cardiometabolic care that used an LLM or other transformer-based generative model to produce responses. Two reviewers independently screened records and extracted data. BCTs were coded using the Behaviour Change Technique Taxonomy v1, and delivery mechanisms for selected self-management BCTs were classified as static (Level 0), rule-based or templated (Level 1), or generative or context-aware (Level 2). Methodological quality of empirical studies was appraised with the Mixed Methods Appraisal Tool, and a study-specific checklist informed by artificial intelligence trial reporting guidance was used to assess implementation reporting transparency.

Results:

Thirty-eight studies met the inclusion criteria. Of these, 21 were classified as empirical human-participant studies. The included studies were concentrated in 2024–2025, reflecting a rapidly expanding but still early-stage literature. Instruction on how to perform behaviour was identified in 30/38 studies (79%), information about health consequences in 27/38 (71%), and feedback and monitoring techniques in 19/38 (50%). Most agents were positioned as educators or coaches targeting type 2 diabetes, obesity, or related cardiometabolic risk, and the technological landscape was dominated by GPT-family models embedded in hybrid architectures with retrieval-augmented generation or rule-based components. Generative outputs were used mainly for tailored explanations, risk information, and socioemotional responses, whereas self-monitoring, reminders, and other structured interactions were more often delivered through rule-based or mixed-mode mechanisms. Among implementation-reporting domains, only 13/38 studies (34%) fully reported prompts or system messages and 16/38 (42%) fully reported safety or oversight mechanisms. User evaluations generally reported good usability and perceived helpfulness, but reported behavioural or physiological outcomes were sparse and were typically limited to pilot, short-term, or single-case designs.

Conclusions:

LLM-driven conversational agents for cardiometabolic care are proliferating but remain an early-stage, methodologically heterogeneous field. Current systems primarily use LLMs as educational and explanatory layers with “synthetic empathy” on top of rule-based data capture and safety functions, and behaviour change content is dominated by information provision and simple feedback. More rigorous comparative studies with longer follow-up are needed before firm conclusions can be drawn about sustained behavioural or clinical benefit. Clinical Trial: OSF Registries JW8VZ; https://osf.io/jw8vz