A feasibility and acceptability study of a large language model-based conversational agent for brief alcohol intervention among emerging adults

About This Grant

PROJECT SUMMARY/ABSTRACT Emerging adults (EAs; ages 18-29 years) have the highest rates of heavy drinking and lowest levels of treatment engagement of any age group. The public health burden of EA alcohol use is immense, costing the U.S. economy billions of dollars each year. Screening and brief interventions (SBIs) are effective at reducing alcohol consumption and problems among EAs and may help to address the treatment gap, yet widespread implementation remains limited due to resource constraints. There is a critical need for innovative, scalable, and cost-effective strategies to increase SBI utilization among EAs engaged in hazardous or harmful drinking. Digital technologies hold promise for addressing the SBI implementation gap given cost-effectiveness and widespread digital technology use among EAs. While digital alcohol interventions have produced small but significant effects on reducing alcohol use and problems, their limited personalization and lack of adaptability may hinder engagement, and addressing this issue may increase the impact of digital efforts. Generative artificial intelligence, namely large language model (LLM)-based conversational agents (e.g., ChatGPT) have the potential to revolutionize automated digital alcohol interventions by allowing for continuous, personalized, and adaptive engagement that is more consistent with a human interaction. In this R34 proposal, we will develop, validate, and conduct an open trial of a LLM-based conversational agent (CA)-delivered brief intervention designed to reduce alcohol use and problems among EAs to establish preliminary feasibility and acceptability. To develop the augmented LLM, we will use instruction fine-tuning to enhance conversational abilities within the context of brief interventions based on high-fidelity recordings of sessions from prior clinical trials and simulated patient-provider interactions. A retrieval augmented generation system will be developed to ensure the model delivers accurate information. The augmented LLM will be incorporated into a CA interface hosted on a web application. To validate the CA’s capability for delivering brief alcohol interventions, we will enroll patient actors (clinical or counseling psychology PhD students) and assign them clinical vignettes depicting a diverse range of EAs engaged in patterns of drinking associated with alcohol use disorder. Patient actors will engage in two randomly ordered online text-based brief intervention sessions for each vignette (one with the CA and one with a human clinician). Blinded dialogues from sessions will be presented to experts and evaluated for treatment fidelity. To maximize and measure initial feasibility and acceptability of the intervention, we will conduct semi-structured interviews (n=20) and an open trial (n=20) with EAs engaged in hazardous drinking. Our primary hypothesis is that the intervention will be deemed feasible and acceptable. We also hypothesize participants will report significant reductions in alcohol use and problems at 1-month follow-up. The intervention developed in this study, if shown to be efficacious in future studies, could improve and increase utilization of alcohol SBIs among EAs, an objective outlined in NIAAA’s strategic plan.