Google says its research medical AI, AMIE, matched primary care doctors in a live video-consultation test. Fifteen trained actors portrayed patients with conditions spanning five body systems, and clinical evaluators rated the system on par with physicians on several core measures.
Real-patient studies must still follow before any conclusions about clinical use can be drawn, Google cautions. The current results come from a randomized evaluation with a specific structure.
AMIE does not hand a single model full control of a consultation. Instead, three agents divide the work: a talker agent carries the conversation with the patient, a planner agent works in the background updating differential diagnoses and management plans and flagging missing information, and a perception agent watches the video and audio streams for non-verbal signs, physical findings, and auditory signals.
The split exists to manage latency. Deep clinical reasoning takes time, and long pauses can damage rapport, so the design lets the talker respond without waiting for every background process to finish.
The evaluation compared AMIE in real-time video, a text-only version, and ten board-certified primary care physicians using the same video interface. Every consultation was later reviewed by a separate panel of 20 experienced physicians working from established rubrics. On history-taking thoroughness, diagnostic accuracy, management appropriateness, and communication quality, the ratings put AMIE on par with the physicians, with the video system scoring higher on eliciting physical signs and guiding actors through virtual examination maneuvers.
