Knowing how far along a pregnancy is underpins almost everything in prenatal care, from screening schedules to decisions about delivery. Yet two-thirds of people worldwide lack access to diagnostic imaging such as ultrasound, and the World Health Organization estimates more than 250,000 preventable deaths a year from pregnancy complications. Google Research has published a prospective study suggesting that health workers with no sonography training, using a cheap handheld probe and an AI model on a phone, can date a pregnancy about as accurately as experts.
The method relies on blind sweeps: instead of hunting for precise views of the fetus, the operator glides the probe across the abdomen in a fixed pattern, and the model does the interpretation. The study, published in JAMA Network Open with Jacaranda Health and Northwestern University, tested 385 participants between 16 and 36 weeks of pregnancy in Chicago and Nairobi, with novice operators and a Clarius C3 probe the model had not been trained on. It had originally learned from scans in North Carolina and Zambia using other probes. The team adapted it with a small pilot set of 120 Chicago participants, about 6% of the original training data, and used no Nairobi data at all, so that site served as an independent test.
The model's average error was 4.18 days, against 4.48 days for expert sonographers using standard measurements, enough to meet the study's test for non-inferiority. It reached 4.10 days in Chicago and 4.27 in Nairobi, and did better than standard formulas for unusually large fetuses. Running in real clinics also produced practical lessons, such as the value of swappable batteries between patients, and the importance of confidence estimates that reject poor sweeps and show uncertainty to the clinician.
This is one of the more useful medical AI results of the year because it targets a basic gap rather than a niche diagnosis, and because the test was prospective, live and partly external. It fits Google's broader push, alongside its tuberculosis screening work, to move health models out of retrospective datasets and into clinics.
The limits are clear in the study's design. Both sites were urban hospitals, not the rural posts where access is worst. The model still needed retuning for a new probe, which any wider rollout would have to repeat. It remains a research system rather than an approved product, and Google gives no timeline for regulatory clearance or deployment. A few days of accuracy on average also says little about rare cases where an error matters most.