Google outlines broader AMIE trials as medical AI moves into patient intake

Google used an October 20 research update to explain how its AMIE medical AI project could change the work surrounding a doctor’s appointment, with broader controlled studies and richer forms of interaction now on the research agenda. Scientists Anil Palepu and Mike Schaekermann described a system that gathers a patient’s history before a visit, then passes a concise account to the clinician. AMIE remains a research project.

The aim is to make the initial exchange more responsive than a fixed intake form. A patient describing back pain, for example, can face follow-up questions about its location, character and onset. In the workflow described by the researchers, this information helps the doctor begin a more focused discussion. The clinician still evaluates the symptoms and works with the patient on treatment decisions; gathering a history is only one part of that encounter.

The update follows results Google announced on October 8 from a study at Beth Israel Deaconess Medical Center’s ambulatory primary care clinic. Ninety-eight patients interacted with AMIE before urgent care appointments. Physicians supervised the conversations in real time, and none required interruption under the study’s predefined safety criteria. That is evidence from a monitored clinical setting, rather than a demonstration that an unsupervised service would behave safely across all patients or conditions.

Google said clinicians found the summaries useful in preparing for visits in 75% of cases and reported that the information influenced their approach to care in more than half. AMIE’s differential diagnoses included the doctors’ final diagnoses 90% of the time. A differential diagnosis is a set of possible explanations: that figure should not be read as a claim that the system independently delivered the correct final diagnosis in nine out of ten consultations.

A separate strand of AMIE research concerns video. In August, Google described a system built with Gemini and Project Astra that combines multiple agents to interpret visual and audio cues and guide virtual examinations. Its randomized evaluation used simulated consultations with patient actors and primary care physicians. Clinical evaluators rated several capabilities favorably, but those results came from a different study design and cannot simply be substituted for outcomes in everyday medical practice.

The project has also examined care over multiple encounters. Research published in Nature in June used one agent for patient dialogue and another for management reasoning, drawing on clinical guidelines and drug formularies. Specialist physicians compared AMIE with 21 primary care doctors in a blinded study involving patient actors. This work addressed management reasoning and the precision of proposed plans, broadening the research beyond a single diagnostic conversation without establishing routine clinical readiness.

The earlier clinic study put the AI before an appointment, so its role was bounded by the next step: a physician reading the summary and seeing the patient. The video and management experiments asked different questions using patient actors. Those boundaries leave an unresolved gap between evaluating components of a consultation and measuring the effects of an integrated service across a larger clinical population.

In the October 20 interview, the researchers said the next work would involve more patients across the United States and controlled comparisons, alongside further exploration of spoken and video interaction. The practical question is whether earlier information gathering can consistently improve the subsequent conversation between patient and physician. Answering it will require the larger clinical studies the team is preparing.