A prospective study described by Google Research and published in JAMA Network Open evaluated an AI model that interprets "blind sweep" ultrasound scans performed with low-cost handheld probes by frontline health workers without specialized sonography training. The model was tested live on mobile devices in real clinics across two settings: an urban academic medical center in Chicago and a community clinic network in Nairobi. Across a primary cohort of 385 participants between 16 and 36 weeks of pregnancy, the AI system was non-inferior to expert sonographers at estimating gestational age, with a combined mean absolute error of 4.18 days versus 4.48 days for the standard expert approach.
The study was designed to test generalizability. The original model trained on data from North Carolina and Zambia was minimally fine-tuned using scans from 120 participants in Chicago, then prospectively tested on 192 participants in Chicago and 193 in Nairobi. Because no Nairobi data was used for fine-tuning, that site served as an independent external test. The evaluation also included multiple-gestation pregnancies, which add tracking and measurement complexity for both human operators and automated systems.
The broader motivation is access: the WHO estimates that more than 250,000 preventable deaths occur each year from pregnancy-related complications, and roughly two-thirds of people globally lack access to diagnostic imaging. Blind-sweep ultrasound paired with AI interpretation could allow minimally trained operators to expand prenatal screening coverage, though this is based on a single source and the article does not discuss independent replication or limitations in detail.