Google's research medical artificial intelligence system, AMIE, has demonstrated the ability to conduct real-time clinical video consultations at an expert level, according to a new study published by Google Research. This first-of-its-kind demonstration involved simulated consultations where AMIE performed comparably to primary care physicians (PCPs) across key clinical areas, including history-taking thoroughness, diagnostic accuracy, management appropriateness, and communication quality.
The advanced version of AMIE, built on Google's Gemini and Project Astra technologies, is designed to perceive non-verbal clinical cues, guide patients through virtual physical examinations, and perform diagnostic reasoning synchronously. In a randomized study involving 100 scenarios and 300 live consultations with 30 board-certified PCPs acting as evaluators, AMIE's video consultation capabilities were assessed. The results indicated that AMIE not only matched human physicians in overall clinical performance but also showed a particular strength in eliciting physical signs and guiding virtual examination maneuvers, scoring significantly higher than PCPs in these aspects.
Beyond clinical evaluations, patient actors participating in the study expressed a strong preference for the video consultation format over text-based interactions. They reported that the video interface was easier to use and more effective for conveying health concerns. Patient actors also rated AMIE favorably for empathy, rapport, and confidence in care when compared to both PCPs and AMIE's text-only version.
This development builds upon previous research into AMIE. Earlier iterations of the system demonstrated expert-level performance in text-based diagnostic dialogues and served as a differential diagnosis aid for clinicians. Google Research had previously extended AMIE's abilities to specialist-level evaluations in fields such as oncology, cardiology, and ophthalmology, as well as multimodal diagnostic reasoning using images and clinical documents in simulated settings. Furthermore, a study published in Nature indicated that AMIE could match primary care doctors in simulated, multi-visit disease-management reasoning and outperformed them in areas such as plan preciseness and guideline alignment.
Google has also been exploring AMIE's application in real-world clinical settings. A prospective feasibility study conducted in partnership with Beth Israel Deaconess Medical Center examined AMIE's ability to gather patient history before ambulatory primary care visits. In this study, AMIE's text-based interactions were overseen by a physician, and the system was found to be feasible, safe, and well-received by both patients and clinicians, with no safety interventions required by human supervisors.
The ongoing research into AMIE aims to enhance access to medical expertise and support physicians by automating certain aspects of patient interaction, allowing them to focus on more complex or personal elements of care. While these advancements show significant promise, the system is still considered experimental and has not yet been tested in actual clinical care for patients. Future work will continue to focus on translating these research advances towards clinical practice through frameworks for physician-centered oversight and further real-world clinical studies.