Intelligent chatbot for online support groups to treat tobacco addiction.

About This Grant

PROJECT ABSTRACT We seek to refine, field test, and deploy an intelligent AI-based chatbot in small, peer-to-peer mobile support groups for smoking cessation. The chatbot will complement the human support group members by responding to posts when no human can. We expect the chatbot to improve engagement by ensuring no post goes unanswered. By developing and testing this chatbot, we hope to breathe new life into research on mobile health support groups which have been challenged by low engagement. We have already built a chatbot using a state-of-the-art open-access LLM (large language model) and set it up on a local, dedicated, secure server. The LLM does not store data on the cloud and all posts are encrypted before server storage. We have trained our chatbot on 77400 posts from our past mobile support groups to accurately comprehend the 25 most common post types. We have developed 25 response libraries for the chatbot that contain over 1k responses developed from knowledge bases, e.g., it provides support for quitting, assists with study-provided NRT, and advises on coping with cravings and stress. Our chatbot intervention is based on the Supportive Accountability Model of Mobile Health. We expect the peers in the group to provide legitimate information that is relevant, trustworthy, and expert, and to form social bonds by being caring, nonjudgmental, and timely, increasing accountability and adherence. By adding a chatbot, we hope to provide additional legitimate information and social bonding. Aim 1 is to refine the chatbot using human-centered design methods. Before putting the chatbot into groups, we want to ensure it can communicate 1:1. We will recruit 4 groups of 5 smokers (N=20), and ask each person to interact with the chatbot, using all 25 post types. They will think aloud in the presence of trained staff to identify and solve usage problems. The sessions will be taped, a survey conducted, and usability metrics assessed. Aim 2 to conduct a one-armed field trial to assess the chatbot’s feasibility and acceptability within human support groups. We will recruit smokers in 2 groups of 10 (N=20). Participants will be placed in a mobile group with the chatbot and asked to support each other in quitting for 2 weeks. We will download and analyze posts and conduct exit interviews. Aim 3 is to conduct a pilot RCT of chatbot efficacy. We will recruit 4 cohorts with 30 smokers per cohort, randomizing 15 smokers to intervention arm (support group with chatbot) and 15 to control arm (support group without chatbot). We will measure primary and secondary engagement outcomes and will have adequate power to compare intervention vs. control. We will also measure intervention-end bioconfirmed abstinence to power a larger RCT.