Conversational AI (AMIE) in Urgent Care: Feasibility and Early Findings

A recent report describes a prospective, single-arm feasibility study that placed Google’s conversational diagnostic AI, the Articulate Medical Intelligence Explorer (AMIE), into an urgent-care workflow and compared its differential diagnoses and proposed management plans with those of primary care providers (PCPs) using blinded assessment.

The piece frames the evaluation around previsit history collection, supervised patient–AI interactions, and a retrospective comparison of diagnostic reasoning and proposed management. The study involved 100 adult patients scheduled for urgent-care visits, with the system used before clinician contact rather than during the encounter. Overall, the work is presented as a feasibility evaluation designed to assess safety, user experience, and clinical reasoning performance in a real-world clinical workflow.

The operational setup described in the article centers on previsit intake. Adults scheduled for urgent-care appointments completed secure, text-based chats with AMIE up to five days before their visit. During the interaction, the system gathered clinical history and generated potential diagnoses intended for patients to discuss with their provider during the upcoming appointment. Patients also reported on their experience after the interaction.

Safety oversight in the study is described as active monitoring during the patient–AI conversations. Human safety supervisors monitored all patient–AMIE interactions in real time, with predefined criteria allowing intervention if safety concerns arose. In the reported study, supervisors did not need to stop any consultations.

For comparative evaluation, the article describes a blinded assessment conducted eight weeks after the clinical visits, using chart review to compare outputs from AMIE and from primary care providers against the finalized clinical diagnosis documented in the medical record after the encounter. In the reported results, blinded evaluators found no statistically significant difference in overall differential diagnosis quality (p = 0.6) and reported that management plan appropriateness (p = 0.1) and safety (p = 1.0) also did not significantly differ between AMIE and PCPs. However, PCPs outperformed AMIE in the practicality (p = 0.003) and cost-effectiveness (p = 0.004) of management plans.

The study also reports diagnostic accuracy metrics for the system. AMIE’s differential diagnosis included the final diagnosis documented in the chart in 90% of cases, with the correct diagnosis appearing within the top three suggestions in 75% of cases.

Patient attitudes were also evaluated. Survey results showed high patient satisfaction with the interaction, and attitudes toward medical AI improved after the chat (p < 0.001). Clinicians also reported that AMIE’s output was useful and had a positive impact on their preparedness for the patient encounter.

The report concludes that conversational AI demonstrated initial feasibility, safety, and user acceptance within a real-world clinical workflow, while emphasizing that further research will be needed before broader clinical implementation.

Key Takeaways:

• Research describes a supervised, previsit urgent-care history-taking workflow using a conversational diagnostic AI (AMIE), with real-time human safety monitoring during patient–AI interactions.
• On blinded assessment against finalized clinical diagnoses, the study reports similar overall differential diagnosis quality and management-plan safety between AMIE and primary care providers, with PCPs outperforming the AI on practicality and cost-effectiveness.
• AMIE’s differential diagnosis included the final chart diagnosis in 90% of cases with 75% top-three accuracy, and patients reported high satisfaction with improved attitudes toward medical AI after interacting with the system.