Research AI dates pregnancies from simple ultrasound sweeps

Picture a health worker with no sonography training sliding a cheap, portable probe across a pregnant woman's belly in a fixed pattern, and an AI model reading the result to estimate how far along the pregnancy is. That is the scenario a Google Research team tested in a prospective study published in JAMA Network Open, with the system running live on mobile devices in real clinics.

The stakes are concrete. The World Health Organization estimates that more than 250,000 preventable deaths occur each year from pregnancy-related complications, and two-thirds of people worldwide lack access to diagnostic imaging like ultrasound. The clinical standard for dating a pregnancy depends on an expert who locates several cross-sections of the fetus and takes precise measurements, a task that takes years of training and varies from one operator to another. As the OB-GYN Nichole Young-Lin puts it, “All of our screening, a lot of our treatments, a lot of our indications for delivery are all based on this master timeline.”

The model was first trained on ultrasound data collected in North Carolina and Zambia, using several commercial probes. For the new study, it was paired with a different handheld probe, the Clarius C3. Adapting it required a small pilot set of 120 participants from Chicago, about 6% of the original training data, to handle the new hardware and to predict gestational age beyond 36 weeks. The Nairobi site, a community clinic network, received no fine-tuning at all, so its results count as an independent external test. The study also included multiple pregnancies, which are harder to measure for human operators and automated systems alike.

In the primary cohort of 385 participants between 16 and 36 weeks, the AI's mean absolute error, meaning the average gap in days between its estimate and the reference, was 4.18 days across both sites. Expert sonographers scored 4.48 days. In Chicago the adapted model reached 4.10 days, and in Nairobi, with no local optimization, 4.27 days. The authors call this non-inferiority: the system matches the expert standard, it does not beat it. In an exploratory analysis, the model also outperformed traditional biometry formulas for fetuses above the 90th percentile in weight. The study is the developers' own, and its results still need to be reproduced in other settings.

For now, the immediate beneficiaries are the clinics where expert sonographers are rare, and the pregnant women who might otherwise go through a pregnancy without a single ultrasound. A tool that needs only a few days of training and a cheap probe could widen prenatal screening there, but only after further validation and a deployment that the study does not describe. Google Research gives no timetable for that step.