Google highlights blind-sweep ultrasound AI study for prenatal care

Google on October 6 highlighted research conducted with Northwestern Medicine in Chicago and Jacaranda Health in Nairobi, Kenya, in which an AI model estimated gestational age from simple prenatal ultrasound sweeps. The published study found its estimates were comparable to a clinical standard in the patients assessed. The finding could matter where trained ultrasound staff are scarce, although the study does not establish that the method improves birth outcomes or can be used routinely across health systems.

In an interview on Google's blog, researcher Angelica Willis described the technique as a way for healthcare workers without previous ultrasound experience to collect images using a predefined sweep across the abdomen. A model analyzes the resulting video and estimates how far a pregnancy has progressed. Google says the system can also indicate fetal position and process scans on the device without Wi-Fi. The published study's primary evaluation, however, concerned gestational-age estimates.

What the Chicago and Nairobi study measured

The prospective study, published online in JAMA Network Open on July 9, tested a mobile AI application using scans collected by novice operators at the two clinical institutions. The wider cohort included 2,043 participants. Its primary evaluation involved 385 people, with 192 in Chicago and 193 in Nairobi, whose pregnancies were between 16 and 36 weeks. Those figures matter when interpreting Google's broader description of roughly 1,000 mothers at each site: the headline accuracy result came from the smaller evaluation group.

In that group, the adapted model's mean absolute error was 4.2 days, against 4.5 days for the clinical standard. The study reported that the model met its prespecified noninferiority test, which allowed a margin of one day. Its mean absolute error was 4.1 days in Chicago and 4.3 days in Nairobi. These are measures of estimation error in the study, rather than evidence that patients subsequently received better care.

The researchers used protocolized sweeps instead of relying on an operator to capture the particular measurements required in a conventional examination. The published evaluation also used a different probe and new clinical settings to test whether the approach generalized beyond earlier work. Its reported validation focused on gestational age; identifying other pregnancy findings at the point of care was described as future work.

Why an earlier pregnancy estimate matters

Google obstetrician-gynecologist Nichole Young-Lin said knowing gestational age helps clinicians time screening and care and prepare for delivery. She also identified equipment costs, difficulties repairing machines, electricity needs and a shortage of sonographers as obstacles to ultrasound access. Smaller handheld devices may ease some equipment barriers, but they do not by themselves remove the need for people able to perform and interpret scans.

The World Health Organization recommends that one imaging ultrasound before 24 weeks of pregnancy should be available to all pregnant women. Its recommendation covers estimating gestational age, detecting fetal anomalies and multiple pregnancies, and improving the pregnancy experience. WHO's obstetric device profile notes that available equipment ranges from conventional cart-based systems to handheld and other portable devices. The blind-sweep study examined pregnancies from 16 through 36 weeks, so its participants were not confined to the period covered by WHO's early-ultrasound recommendation.

What remains to be established before wider use

The study measured diagnostic performance at two urban clinical institutions. It did not demonstrate that introducing the application into routine care would increase access, change clinical decisions or improve maternal or newborn health. WHO has also cautioned that evidence for benefits from early routine ultrasound comes from trials in high-income countries, while the wider health-system effects of introducing antenatal ultrasound in rural, low-resource settings remain unproven.

Google says its model can process scans on the device without Wi-Fi or an electricity supply. That description is relevant to clinics with unreliable connections or power, but the diagnostic comparison does not establish how the system would perform during sustained routine use in such conditions. Questions also remain about how well results would carry over to other populations, operators and care pathways.

The findings advance a project that Northwestern described in 2022 as being in a data-gathering and algorithm-development stage. The paper says Google funded grants to Jacaranda Health and Northwestern, and Google-employed authors took part in study design, analysis, interpretation and writing; Google was not directly involved in collecting the data. The next question is whether further evaluations can show dependable performance and useful care outcomes beyond the two institutions studied.