Behavior Change Content and Implementation of Large Language Model-Driven Conversational Agents in Cardiometabolic...

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

This scoping review mapped behavior change techniques (BCTs) used in LLM-driven conversational agents for cardiometabolic prevention and management, described how these techniques are delivered across static, rule-based, and generative mechanisms, examined LLM design, personalization, and safety reporting, and summarized user experience and behavioral or clinical outcomes.

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 from January 1, 2020, to November 30, 2025. The final search was run on March 25, 2026, using this publication-date limit. Eligible studies reported a patient-facing text- or voice-based cardiometabolic conversational agent using an LLM or other transformer-based generative model. Two reviewers independently screened records and extracted data. BCTs were coded using the Behavior Change Technique Taxonomy v1; selected self-management BCTs were classified as static, rule-based or templated, or generative or context-aware. Empirical human-participant- or evaluator-based studies were appraised with the Mixed Methods Appraisal Tool, and a study-specific checklist assessed LLM implementation reporting transparency.

Thirty-eight studies were included; 19 involved empirical human-participant- or evaluator-based assessments, whereas 19 were technical and system-level evaluations, including framework-development, simulated-output, and proof-of-concept studies. Studies were concentrated in 2024-2025. Instruction on how to perform behavior was identified in 30 of 38 (79%) studies, information about health consequences in 27 of 38 (71%) studies, and feedback and monitoring techniques in 19 of 38 (50%) studies. Most agents were positioned as educators or coaches targeting type 2 diabetes, obesity, or related cardiometabolic risk, and GPT-family models embedded in hybrid architectures with retrieval-augmented generation or rule-based components predominated. Generative outputs were used mainly for tailored explanations, risk information, and socioemotional responses, whereas self-monitoring, reminders, and structured interactions were more often rule-based or mixed-mode. Only 13 of 38 (34%) studies fully reported prompts or system messages, and 16 of 38 (42%) studies fully reported safety or oversight mechanisms. User evaluations reported good usability and perceived helpfulness, but behavioral or physiological outcomes were sparse and usually limited to pilot, short-term, or single-case designs.

LLM-driven conversational agents for cardiometabolic care are proliferating but remain early-stage and methodologically heterogeneous. Current systems primarily use LLMs as educational and explanatory layers with "synthetic empathy" over rule-based data capture and safety functions, while behavior change content remains 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 behavioral or clinical benefit.