← Back to Home

Discussing Symptoms with AI Before a Visit: AMIE Study Completes Clinical Testing Under Physician Supervision

Moving from simulated consultations into real outpatient care, Google’s AMIE has provided preliminary evidence of feasibility. No conversations triggered a safety stop during the study, but the performance of its diagnostic lists remains some distance from demonstrating improved patient health outcomes.

By SURL BioNews

Before entering the examination room, could patients use a text conversation to describe their symptoms and medical history more fully, helping clinicians identify the key issues ahead of the visit? Google’s AMIE study, conducted in collaboration with Beth Israel Deaconess Medical Center (BIDMC), brought this idea into real primary care clinics. On October 8, the team announced that the study had been published in *The Lancet*. It provides evidence of clinical feasibility under physician supervision and moves the discussion of medical AI from its ability to answer questions toward its role in actual care workflows.

This was a single-center, single-arm prospective study, with no randomized control group. According to ARISE’s study description, 100 adults completed a text conversation with AMIE, and 98 also completed the subsequent visit. The study preprint states that conversations took place up to five days before primary care visits scheduled for acute symptoms. The system collected medical histories, proposed possible diagnoses, and then provided the conversation transcripts and summaries to the clinicians conducting the visits. The use case here was preparation for outpatient visits, rather than emergency resuscitation.

A physician monitored every conversation in real time and determined whether to stop it according to predefined safety criteria. No conversations triggered a safety stop during the study. However, this result must be understood in the context of those supervision conditions. The preprint records that clinic triage staff had already determined that participants did not require emergency care, and that the study excluded people known to be pregnant and those whose chief complaint concerned mental health. The absence of safety stops therefore does not establish that the system is suitable for all patients or can operate safely without supervision.

Diagnostic evaluation was based on a medical record review eight weeks after the visit. Among the 98 participants who completed both the conversation and the visit, the final diagnosis appeared among AMIE’s seven possible diagnoses in 90% of cases and among its top three in 75%. This reflects whether the system could include the answer within its differential diagnosis. It cannot be interpreted directly as 90% accuracy for a single diagnosis, nor does it mean that patients would achieve better treatment outcomes by acting on AI recommendations.

Blinded review also revealed more nuanced differences: there were no statistically significant differences between AMIE and clinicians in ratings of overall differential diagnosis quality or the appropriateness and safety of management plans. However, clinicians’ plans performed better in practicality and cost-effectiveness. The corresponding p-values reported in the preprint were 0.003 and 0.004. A lack of statistical significance does not establish equivalence, and these management plan ratings were not comparisons of actual treatment effectiveness or healthcare spending.

For outpatient workflows, the more direct signal came from user experience. Google said that clinicians found the AI summaries helpful for visit preparation in 75% of cases. The preprint also recorded an improvement in patients’ attitudes toward AI after the interaction. However, perceived helpfulness and actual time savings, fewer missed diagnoses, or reduced workload are distinct questions, and this study has not yet established evidence for the latter outcomes.

The study’s preprint was already publicly available in March 2026. The latest development is the team’s announcement of formal journal publication, rather than a new set of clinical trial results. It provides a testable starting point for AI consultations before outpatient visits. The next step requires larger, multicenter studies with control groups to determine whether the benefits extend to different patients and care settings, and to clarify the level of supervision needed in actual deployment.

References

  1. Google
  2. ARISE
  3. arXiv