Background

Generative AI powered Conversational Agents (CAs), e.g. ChatGPT, are increasingly used to automate elements of patient healthcare. We developed a novel CA fine-tuned for bariatric care and assessed its ability to provide patients safe and accurate support. Additionally, we explored the role of human oversight in screening messages for indications of clinical issues or weight loss concerns.

Methods

Consecutive text-based conversations between patients and a novel CA were reviewed. A predefined framework (FAST), encompassing 4 domains, determined quality: (1) Fidelity (advice practical, implementable, and embedded in behavioral science); (2) Accuracy (scientific and clinical validity of responses); (3) Safety (responds appropriately to risk); (4) Tone (empathetic, non-judgmental, and collaborative approach). All patient messages were screened and relayed to healthcare providers if they indicated a potential clinical or weight loss concern.

Results

The analysis comprised 1,197 conversations with patients from 30 bariatric clinics in 9 countries. Average percentage of CA interactions rated acceptable for fidelity, accuracy, safety, and tone were 82%, 74%, 83% and 94%, respectively. Notably, 6.9% of patient messages were relayed to providers. Of these, 71% were due to potential clinical concern, while 29% were due to a potential weight-loss concern. Distribution across treatment types were swallowable intragastric balloon (IGB) 55%; endoscopic IGB 31%; and bariatric surgery 14%.

Conclusions

Based on these findings, CAs demonstrate significant potential to safely enhance patient support for bariatric treatments. Further research will evaluate an additional layer of AI monitoring that automates the identification of concerning messages, thereby reducing the necessity for expert human review.