Google's AMIE medical AI system performed on par with primary care physicians in a study using trained actors for simulated clinical video consultations.

Google has developed a medical AI system called AMIE that conducts synchronous video consultations. In a study involving 15 trained actors portraying various medical conditions, the system received clinical evaluator ratings comparable to those of primary care physicians across several core metrics, including diagnostic accuracy and communication quality.

Multi-agent architecture design

AMIE utilizes an asynchronous multi-agent architecture to manage the complexities of a medical consultation. A 'talker' agent handles spoken interaction to maintain conversational flow, while a 'planner' agent runs in the background to update diagnoses and management plans. A third 'perception' agent continuously reviews audio and video streams to identify non-verbal signs and physical findings, integrating them into the clinical context.

Study methodology and evaluation

The research involved a multi-arm randomized study where AMIE completed real-time video consultations alongside a text-only version and a group of 10 board-certified primary care physicians. An independent panel of 20 experienced physicians evaluated the consultations using established clinical rubrics. Evaluators found that AMIE performed as well as the physicians in history-taking, diagnostic accuracy, and management appropriateness.

Limitations and future research

Google noted several limitations, including the use of professional actors who cannot fully replicate the variability of real-world patient encounters. The study also excluded medical presentations that were difficult for actors to portray. Google emphasized that the current findings do not establish that AMIE can safely diagnose or manage real patients, and further research involving real patients is required.