Integration of a Trained Language Model to Improve Glycemic Control Through Increased Physical Activity: a Fully Digital My Heart Counts Smartphone App Randomized Trial

NCT ID: NCT06596330

Last Updated: 2024-09-19

Basic Information

• Recruitment Status: NOT_YET_RECRUITING
• Clinical Phase: NA
• Total Enrollment: 1000 participants
• Study Classification: INTERVENTIONAL
• Study Start Date: 2025-07-31
• Study Completion Date: 2029-07-31

Brief Summary

Type 2 diabetes (T2D) is one of the most common and fastest growing diseases, affecting 1 in 8 adults (nearly 800 million) worldwide by 2045. Sedentary behavior and increased adiposity are major risk factors for T2D. Cardiovascular disease is the leading cause of death in those with T2D, while diabetic microvascular disease, causing kidney disease, neuropathy, and retinopathy, contributes to T2D morbidity.

Physical activity is one of the most potent therapies in preventing/treating T2D and its complications. Mean daily steps is a proxy for physical activity, with even modest improvements in step count (i.e., +500 steps) associated with decreased T2D and mortality. However, adherence to regular physical activity remains low in T2D patients, with short-term decreases in daily step count associated with impaired glycemic control and T2D recurrence.

The investigators have developed an artificial intelligence (AI) language model (similar to ChatGPT), which can automatically generate coaching prompts to encourage physical activity by incorporating an individual's stage of change. The investigators will extend our research using the My Heart Counts (MHC) smartphone app to 1) validate the efficacy of the AI-generated prompts in patients with T2D and 2) perform a longer-term randomized crossover trial using the language model as a social accountability chatbot - encouraging participants to maintain their physical activity changes over months. The investigators hypothesize that my AI-assisted coaching prompts will significantly increase 1) mean daily step count by 500 steps in 1,000 adults recruited nationwide over a 7-day period, and 2) improve HbA1c and weight via long-term behavior change over a 24-week intervention period.

Conditions

Type2diabetes

Study Design

• Allocation Method: RANDOMIZED
• Intervention Model: CROSSOVER
• Primary Study Purpose: TREATMENT
• Blinding Strategy: NONE

Study Groups

LLM-generated coaching prompt

Group Type: EXPERIMENTAL

Validation of language model prompts in increasing short-term physical activity

Intervention Type: BEHAVIORAL

Aim 1: In preliminary data, the investigators have pre-trained an open-source language model, LLAMA, with expert-created coaching prompts based on the stages of change model for physical activity. Seven different prompts (for each day of an intervention "week") will be generated, accounting for race/ethnicity, age, gender, and stage of change, to improve personalization. Using the existing MHC app, the investigators will perform a randomized crossover trial on mean daily steps across each intervention. The investigators will compare the interventions of a daily reminder to reach 10,000 steps (a neutral control) and AI-personalized interventions based on an individual's stage of change.

Assessment of long-term changes to physical activity and glycemic control

Intervention Type: BEHAVIORAL

Aim 2: Using social accountability and the trained language model generating personalized coaching interventions, the investigators will conduct a long-term follow-up randomized, unblinded trial. Over a 24-week intervention period, participants will receive either a generic daily reminder to reach 10,000 steps or an AI-generated coaching prompt, with the AI group also being able to "chat" with the language model to ask for advice on maintaining their physical activity. The outcomes of this long-term trial will be change in: 1) daily steps over the intervention period, 2) weight (via HealthKit link to MHC), and 3) HbA1c (as derived from EMR records linked to the HIPAA-compliant MHC app).

10,000 Step Reminder

Group Type: ACTIVE_COMPARATOR

Validation of language model prompts in increasing short-term physical activity

Intervention Type: BEHAVIORAL

Aim 1: In preliminary data, the investigators have pre-trained an open-source language model, LLAMA, with expert-created coaching prompts based on the stages of change model for physical activity. Seven different prompts (for each day of an intervention "week") will be generated, accounting for race/ethnicity, age, gender, and stage of change, to improve personalization. Using the existing MHC app, the investigators will perform a randomized crossover trial on mean daily steps across each intervention. The investigators will compare the interventions of a daily reminder to reach 10,000 steps (a neutral control) and AI-personalized interventions based on an individual's stage of change.

Assessment of long-term changes to physical activity and glycemic control

Intervention Type: BEHAVIORAL

Aim 2: Using social accountability and the trained language model generating personalized coaching interventions, the investigators will conduct a long-term follow-up randomized, unblinded trial. Over a 24-week intervention period, participants will receive either a generic daily reminder to reach 10,000 steps or an AI-generated coaching prompt, with the AI group also being able to "chat" with the language model to ask for advice on maintaining their physical activity. The outcomes of this long-term trial will be change in: 1) daily steps over the intervention period, 2) weight (via HealthKit link to MHC), and 3) HbA1c (as derived from EMR records linked to the HIPAA-compliant MHC app).

Interventions

Validation of language model prompts in increasing short-term physical activity

Aim 1: In preliminary data, the investigators have pre-trained an open-source language model, LLAMA, with expert-created coaching prompts based on the stages of change model for physical activity. Seven different prompts (for each day of an intervention "week") will be generated, accounting for race/ethnicity, age, gender, and stage of change, to improve personalization. Using the existing MHC app, the investigators will perform a randomized crossover trial on mean daily steps across each intervention. The investigators will compare the interventions of a daily reminder to reach 10,000 steps (a neutral control) and AI-personalized interventions based on an individual's stage of change.

Intervention Type: BEHAVIORAL

Assessment of long-term changes to physical activity and glycemic control

Aim 2: Using social accountability and the trained language model generating personalized coaching interventions, the investigators will conduct a long-term follow-up randomized, unblinded trial. Over a 24-week intervention period, participants will receive either a generic daily reminder to reach 10,000 steps or an AI-generated coaching prompt, with the AI group also being able to "chat" with the language model to ask for advice on maintaining their physical activity. The outcomes of this long-term trial will be change in: 1) daily steps over the intervention period, 2) weight (via HealthKit link to MHC), and 3) HbA1c (as derived from EMR records linked to the HIPAA-compliant MHC app).

Intervention Type: BEHAVIORAL

Eligibility Criteria

Inclusion Criteria

• Individuals aged ≥18 years old, with a clinical diagnosis of T2D, able to read and understand English, and who are physically able to walk, will be included in our study

• Minimum Eligible Age: 18 Years
• Eligible Sex: ALL
• Accepts Healthy Volunteers: No

Sponsors

Stanford University

OTHER

Sponsor Role: lead

Responsible Party

Daniel Seung Kim

Fellow in Medicine

Responsibility Role: PRINCIPAL_INVESTIGATOR

Central Contacts

Dan Kim, MD, PhD, MPH

Role: CONTACT

kimds@stanford.edu

650-498-4900

Other Identifiers

76338

Identifier Type: -

Identifier Source: org_study_id