Google's AI Matches Primary Care Doctors in Patient Treatment

Google has published results in the journal Nature showing that its experimental medical AI system AMIE didn't just make diagnoses, but managed patients across multiple visits, adjusting treatment as symptoms and test results changed. In simulated scenarios, the system matched or outperformed a group of 21 primary care physicians.

AMIE, short for Articulate Medical Intelligence Explorer, is a research project from Google Research and Google DeepMind that has been in development for several years as a system for conducting diagnostic conversations with patients. Earlier publications mostly tested one-off consultations, where the AI made an initial diagnosis based on a description of symptoms. This new study goes a step further, examining whether such a system can handle care that unfolds over time, the kind of ongoing management a primary care physician provides to a patient across weeks or months of treatment.

How the system works

AMIE's architecture rests on two cooperating modules. The first, a dialogue agent, talks with the patient, gathers information about how they're feeling, and responds to their questions empathetically. The second, a management reasoning agent, analyzes the gathered data against an extensive base of clinical guidelines and proposes a concrete course of action.

This second module draws on the long context window of Gemini models, processing around 627 clinical guideline documents each time, equivalent to roughly 10.5 million tokens. That lets the system factor in current treatment recommendations on the fly, rather than relying solely on knowledge baked in during training.

Test results

The study used a format similar to the OSCE exam, the standard method for assessing medical students' clinical skills. The 100 patient scenarios spanned seven specialties: cardiology, pulmonology, gynecology and obstetrics, urology, gastroenterology, neurology, and orthopedics. Each scenario unfolded over three visits, during which symptoms, test results, and the response to previously started treatment changed.

Thirty independent specialist physicians rated the quality of the management plans without knowing whether they came from AMIE or from a primary care doctor. In assessing the accuracy of treatment recommendations, AMIE scored around 95-96 percent across successive visits, while the primary care physicians scored 62-67 percent. A similar gap held up in the assessment of how precisely additional tests were ordered.

Making a diagnosis is just the first step in treating a patient. Once the diagnosis is set, the challenge becomes managing the condition over time, tracking symptoms across visits, working through updated guidelines, and fine-tuning medications - Mike Schaekermann, research lead, Google Research

Caveats from the authors

The research team is explicit that the results don't mean the system is ready for real clinical work. The simulation relied on scenarios written for the study, actors playing patients, and text-based communication over short intervals, without electronic health records, prescribing systems, pharmacist involvement, or the complexity typical of everyday medical practice.

Google has not released the code or model weights, citing the risk of the medical tool being used without clinical oversight. The company stresses that AMIE remains a research project, not a product ready for deployment with patients.

What's next

Google is already running two further rounds of verification outside the lab. One is a nationwide virtual care study conducted with clinical partner Included Health, the other a feasibility test carried out jointly with Beth Israel Deaconess Medical Center. Both aim to test whether the system holds up under conditions closer to real practice, with actual patients and real staff workloads.

In parallel, Google Research and Google DeepMind are developing a variant of AMIE capable of conducting real-time video consultations, interpreting visual and audio cues and guiding patients through basic elements of a physical exam.

For Poland's healthcare system, where a shortage of primary care physicians and long waits for specialists have been a persistent problem, findings like these carry practical, if still distant, significance. A system able to consistently apply up-to-date clinical guidelines and prescribe medications with precision could eventually take some administrative and documentation work off doctors' hands, freeing up more time for direct contact with patients. For now, though, as the study's own authors point out, the path from a Nature publication to safe use at the bedside still requires years of regulatory and clinical testing.