Google AI estimates pregnancy stage from simple ultrasound sweeps like experts

In a blind sweep, the operator simply glides a portable probe across the abdomen in a set pattern, without having to find specific views of the fetus. The model was first trained on scans from North Carolina and Zambia, then adapted to a new Clarius handheld probe using data from 120 Chicago participants. It got no tuning at all for Nairobi, yet its average error there was 4.27 days, close to Chicago’s 4.10. The team also picked up practical lessons, such as swappable batteries working better than charging docks between patients.

Why it matters

The WHO estimates more than 250,000 preventable pregnancy-related deaths a year, and two-thirds of people worldwide lack access to imaging like ultrasound. A tool that holds up across hospitals, countries and hardware without retraining is what scaling would need. It remains a research system, and the team says building trust, with safeguards such as rejecting poor-quality scans, will matter as much as accuracy.