Google’s AMIE studied in single-center primary-care trial

AMIE stands for Articulate Medical Intelligence Explorer.[1] In the study, English-speaking adults used it before an urgent primary-care appointment for a single complaint.[1] The system asked about symptoms and medical history, suggested possible diagnoses to discuss with a physician, and generated a clinician summary before the visit.[3] A physician monitored each conversation in real time.[3] The work took place at BIDMC’s ambulatory primary-care clinic.[2] For health teams, this setup is an important boundary: the system was tested as preparation for a clinician visit, not as an unmonitored substitute for one.

Clinical evaluators rated AMIE’s conversations favorably in 87% to 100% of cases across 17 criteria; patients gave favorable ratings in 48% to 96% of cases across 16 criteria.[1] AMIE’s differential diagnosis included the final diagnosis in 90% of cases, and its diagnostic accuracy among its top three possibilities was 75%.[4] The comparison found no significant difference in differential-diagnosis quality, while primary-care providers outperformed AMIE on the practicality and cost-effectiveness of management plans.[4] These findings point to different performance across tasks, rather than a single measure of clinical readiness.

No interaction met the study’s predefined criteria for a safety stop, but safety supervisors noted one hallucination and added clinical information in five interactions.[1] In one case, a physician reported possible harm because a patient may have become anxious after AMIE listed lymphoma as a possible diagnosis.[3] Patients’ attitudes toward AI improved after interacting with AMIE and remained elevated after their physician visit.[1] Providers said the system helped them prepare for 33 of the 44 visits where they reviewed its transcript beforehand, and might have changed their behavior in 25 of those visits.[1] The researchers designed the study to assess feasibility, not health outcomes; Google says larger clinical trials are needed to assess patient-facing AI at scale.[3][2]