Google tests AMIE for clinical video consultations

Details

Google's AMIE (Video) system conducted real-time video medical consultations with 15 trained professional actors portraying conditions across five clinical categories: cardiopulmonary, abdominal, HEENT, neurological/psychiatric, and musculoskeletal. An independent panel of 20 board-certified primary care physicians rated every consultation using established clinical rubrics. AMIE matched physicians across several core measures including history-taking, clinical reasoning, treatment recommendations, and patient-centred communication.\n\nThe architecture separates work into three concurrent agents to handle the fundamental latency problem in clinical AI. A 'talker' agent maintains conversational flow with the patient. A 'planner' agent runs in the background, updating differential diagnoses and identifying missing information as the consultation progresses. A 'perception' agent processes video and audio continuously, identifying non-verbal cues and physical findings. This separation allows the patient-facing component to respond without waiting for the slower reasoning processes — a practical engineering solution.\n\nGoogle is explicit about the study's limitations: all interactions used professional patient actors in controlled scenarios, not real patients with genuine health conditions. The company states clearly that studies involving real patients must follow before any clinical conclusions can be drawn. Automated evaluations confirmed that each individual agent component improved clinical measures when added. This is research-stage work, not a product launch.