ORIGINAL RESEARCH
Winston Liaw, MD, MPH; Hien Van Nguyen, PhD; Tonghui Xu, PhD; Quang Hung Bui; Diana Grair, MD; Alauna Hunt, MPH, MSPCS; Carlos Fuentes, BS; Yahaira Suchil, BS; LaShaune Johnson, PhD; William Elder, PhD; James Bray, PhD; Summer Chavez, DO, MPH, MPM; Kimberly Pilkinton, MD, MPH; Daniella Hernandez, CHW; Omolola Adepoju, PhD
Corresponding Author: Winston Liaw, MD, MPH; Department of Health Systems and Population Health Sciences, Tilman J. Fertitta Family College of Medicine, University of Houston;
Email: wliaw@central.uh.edu
DOI: 10.3122/jabfm.2025.260102R1
Keywords: Artificial Intelligence, Clinical Decision Support Systems, Doctor-Patient Relations, Health Behavior, Natural Language Processing, Primary Health Care
Dates: Submitted: 3/12/2026; Accepted: 5/13/2026
Status: In Press.
BACKGROUND: Behavior counseling using the 5A’s model (assess, advise, agree, assist, and arrange) is effective but inconsistently delivered. In response, we developed Primary Care (PC) Navigator, a tool that integrates audio and video with a large language model (LLM) to generate 5A behavior change plans during visits. Our objectives were to assess its impact on behavior change and the acceptability, tone, and quality of its output.
METHODS: In this pre–post pilot, we used PC Navigator to generate behavior change plans. Patients completed self-determination theory (clinician support, autonomy, and competence) measures before and after visits. Patient and clinician acceptability was assessed using the Technology Acceptance Model. Documentation quality of the LLM was evaluated using the Physician Documentation Quality Instrument. We used Bayesian paired t-tests and sentiment analysis to compare the tone and quality of AI- and clinician-generated plans.
RESULTS: During 18 encounters, PC Navigator was associated with a pre-post increase in one clinician-support item (5.2 vs. 6.3/7; p=0.03), with no changes in autonomy or competence. A majority of patients agreed that the tool was easy to use (66.7%) and useful (72.2%). Clinicians were less favorable. Sentiment analysis demonstrated higher proportions of negative (0.4% vs. 0.2%, p<0.001) and sadness (2.6% vs. 1.1%, p<0.001) words in AI outputs. AI-generated plans were more thorough and comprehensible but less succinct (p<0.001).
CONCLUSIONS: PC Navigator was acceptable to patients, producing behavior change plans comparable in quality to clinicians. Ongoing refinement of the LLM is needed to optimize tone, clinician acceptability, and impact on behavior change.