The Pulse
Google Tests AI Ultrasound With Novice Operators in Chicago and Nairobi
Google Research described a prospective study in which AI estimated pregnancy stage from ultrasound sweeps collected by novice operators in Chicago and Nairobi. In 385 participants, the model’s average error was 4.2 days, compared with 4.5
AI.info Team · Written with AI · fact-checked ·
Google Research said on October 21 that an AI system estimated pregnancy stage from ultrasound scans performed by novice operators in Chicago and Nairobi, with results comparable to expert sonographers. The announcement describes a research study, not a public product launch: the study appeared in JAMA Network Open on July 9, 2026. Its main analysis covered 385 participants between 16 and 36 weeks of pregnancy.
Google’s October post details a July study
The system targets gestational age—the estimate of how far along a pregnancy is—using what researchers call “blind sweeps.” Rather than asking an operator to identify fetal anatomy and capture specific measurements, the protocol has them move a handheld probe across the abdomen in a prescribed sequence. An AI model then analyzes the ultrasound video.
The research team tested the approach at Northwestern Medicine in Chicago and Jacaranda Health’s clinic network in Nairobi. It used a Clarius C3 handheld transabdominal probe, a device not included in the model’s original training data. The model had been trained using scans from North Carolina and Zambia, then adapted using data from 120 Chicago participants.
The model’s error averaged 4.2 days
Across the 385-person evaluation group, the AI model’s mean absolute error was 4.2 days, compared with 4.5 days for estimates based on the standard expert-sonographer procedure. The study found the model met its predefined test for noninferiority, which allowed a one-day difference. In Chicago, the model’s error averaged 4.1 days; in Nairobi, it averaged 4.3 days.
The Nairobi results are notable because the researchers did not use local data to fine-tune the model. That makes the site an external test of whether a system adapted in Chicago could perform in a second setting. The paper’s authors write: “These findings suggest that the AI model for GA estimation, using blind sweep data collected by novice operators, demonstrated generalization to new clinical settings and a different ultrasonography probe.” Angelica Willis et al., JAMA Network Open
Training differed between the two clinics
The operators had no prior ultrasound experience, but their preparation was not identical. In Chicago, two operators received a handout, verbal instructions and a demonstration; in Nairobi, three operators completed six hours of hands-on training. The study recorded a higher rate of rejected scans in Chicago than in Nairobi, though those differences do not establish that training alone caused the gap.
The study also tested a narrow task: estimating gestational age, not screening for pregnancy complications or replacing a full diagnostic ultrasound. During data collection, the model’s results were hidden from operators to avoid influencing clinical care. The researchers say further testing is needed across more ultrasound hardware; the paper’s evaluation covered one new probe manufacturer and two clinical settings.
Researchers identify practical hurdles to use
Google’s account of the study points to day-to-day constraints as well as model performance. Staff needed to keep devices charged between consecutive examinations, and the research team found swappable batteries more useful than plug-in docking stations. Google also says future versions should provide confidence estimates that can flag poor-quality scans and communicate uncertainty to clinicians.
The results support continued study of AI-assisted ultrasound as a way for less-specialized staff to collect pregnancy data. They do not establish that the system is ready for routine care or that it can take on broader obstetric diagnosis. For now, the evidence is specific: a research app, a prescribed sweep protocol, a single newly tested probe and gestational-age estimates from 385 participants in Chicago and Nairobi.