Effectiveness of an LLM-integrated Chatbot Intervention for Physical Activity Habit Formation in Prehypertensive Adults: A 12-Month Randomized Controlled Trial Protocol

Technical Summary: Effectiveness of an LLM-Integrated Chatbot Intervention for Physical Activity Habit Formation in Prehypertensive Adults

Problem Statement
Prehypertension is a critical risk factor for hypertension and cardiovascular disease, affecting a significant portion of the adult population. While physical activity (PA) is a proven non-pharmacological intervention for managing blood pressure, long-term adherence remains a major challenge. Existing interventions often rely on structured exercise programs or standardized advice that target reflective processes (e.g., knowledge, intention, self-efficacy) but fail to support the reflexive processes necessary for habit formation (e.g., automatic cue-behavior associations, context stability). Furthermore, while digital health tools like wearables and chatbots exist, there is limited evidence on the effectiveness of integrating Large Language Models (LLMs) to provide adaptive, context-sensitive, and continuous support for sustained PA behavior change in prehypertensive adults. Current LLM-based health studies are often short-term, focusing on feasibility rather than long-term behavioral outcomes or the mechanisms of habit formation.

Methodology
This paper outlines the protocol for a 12-month, two-arm, parallel, pragmatic Randomized Controlled Trial (RCT) with an embedded 3-month intensive longitudinal study. The trial is registered with the Chinese Clinical Trials Registry (ChiCTR2400085073).

• Participants: 118 adults (aged 18–60) with prehypertension (SBP 120–139 mmHg or DBP 80–89 mmHg) and low physical activity levels will be recruited from two large technology companies in Beijing and via social media.
• Design:
  • Intervention Group: Participants receive a Huawei Watch D smartwatch and access to HabitBot, an LLM-integrated chatbot delivered via a WeChat mini-program. HabitBot is grounded in the Health Action Process Approach (HAPA) and Habit Formation Theory. It utilizes a user-centered design to deliver personalized PA prescriptions, health advice, and motivational feedback. The system incorporates participant profiles, real-time wearable data (steps, heart rate, BP), and conversational history to generate adaptive responses. It supports behavior change techniques including goal setting, action planning, problem-solving, and self-monitoring.
  • Control Group: Participants receive the same Huawei Watch D smartwatch, a one-time face-to-face tailored PA prescription from a nurse (based on ACSM guidelines), and access to a standard health app (CoHealth) providing general education on PA, diet, and sleep.
• Randomization: Randomization occurs at the cluster level for company-recruited participants (to manage logistical constraints) and at the individual level for social media recruits.
• Outcomes:
  • Primary: Average daily step count (measured objectively via the smartwatch) and physical activity levels (via IPAQ).
  • Secondary: Blood pressure, body composition, and psychological factors related to reflective (HAPA-based: intention, self-efficacy, planning) and reflexive (SRBAI-based: habit strength, context stability, intrinsic reward) pathways.
  • Process Measures: Usability (Chatbot Usability Scale), engagement, and adherence.
• Data Collection: Assessments occur at baseline, 3, 6, and 12 months. An intensive Ecological Momentary Assessment (EMA) is conducted daily during the first 3 months to capture temporal changes in habit formation mechanisms.
• Analysis: Intention-to-treat analysis using mixed-effects models to account for repeated measures and clustering. Mediation analyses will explore how changes in reflective and reflexive pathways influence PA outcomes.

Key Contributions
The paper proposes a novel framework for addressing the gap in long-term PA maintenance by integrating three specific elements:

1. LLM-Driven Adaptivity: Unlike rule-based chatbots, HabitBot uses an LLM to handle multi-turn dialogues, context understanding, and dynamic problem-solving, allowing it to adjust to user barriers and changing health status in real-time.
2. Dual-Pathway Focus: The intervention explicitly targets both reflective (motivational) and reflexive (habitual) processes, aiming to transition users from intentional exercise to automatic habit.
3. Mechanistic Longitudinal Design: By embedding a 3-month intensive daily assessment within a 12-month trial, the study aims to elucidate the temporal mechanisms of habit formation, rather than just measuring end-state outcomes.

Results
As this document is a study protocol, it does not report empirical results, statistical findings, or efficacy data. The "Results" section of the paper outlines the anticipated contributions: the study is expected to provide insights into the effectiveness of HabitBot in promoting PA habits and improving health among prehypertensive adults, as well as exploring the mechanisms underlying habit formation through longitudinal data analysis.

Significance and Claims
The authors claim that this study addresses three critical gaps in current literature:

1. The lack of interventions that integrate quantified PA guidance, objective wearable monitoring, and continuous behavioral support into a single system.
2. The over-reliance on reflective determinants in PA interventions, with insufficient attention to reflexive habit formation processes.
3. The scarcity of rigorous, long-term evidence regarding the efficacy of LLM-integrated chatbots for sustained behavior change.

The study posits that if successful, the findings will help shape future LLM-integrated interventions for chronic disease management. By examining both reflective and reflexive pathways, the research aims to advance digital health strategies for reducing cardiovascular risk, moving beyond simple motivation to support the automatic initiation of healthy behaviors in daily life. The authors maintain a modest tone, acknowledging limitations such as the inability to blind participants to the intervention, potential measurement errors in wearable data, and the specific demographic (Beijing-based tech workers) which may limit generalizability.