Google Research published the results of a prospective multicenter study of an AI system that estimates gestational age from 'blind' ultrasound scans in JAMA Network Open. The blind sweep protocol was performed by healthcare workers without sonographer training on a portable Clarius C3 sensor, while the model calculated the gestational age. In a cohort of 385 participants in Chicago and Nairobi, the AI's mean error was 4.18 days compared to 4.48 days for experts, falling within the pre-specified non-inferiority threshold of one day, with the model not being fine-tuned for the local population in Nairobi. The study confirms that a combination of a simple scanning protocol, a low-cost sensor, and AI can provide expert-level accuracy where qualified obstetric ultrasound is unavailable.

image
image

What happened

Google Research published the results of a prospective multicenter study of an AI system for gestational age assessment in the journal JAMA Network Open (2026, DOI 10.1001/jamanetworkopen.2026.22484). The main cohort consisted of 385 participants with gestational ages of 16–36 weeks: 192 at Northwestern Medicine in Chicago and 193 at Jacaranda Health in Nairobi. Scans were performed by non-specialists using the blind sweep protocol on a portable Clarius C3 sensor, and the model determined the gestational age from the resulting recording. The AI's mean absolute error was 4.18 days compared to 4.48 days for expert sonographers; in Chicago, the model's error was 4.10 days, and in Nairobi, it was 4.27 days, with no local adaptation for the Kenyan population. The difference from experts remained within the pre-specified non-inferiority threshold of one day. In parallel, Google described the study in its blog.

Context

Gestational age is a fundamental parameter in pregnancy management: accurate dating affects the examination plan, complication recognition, and delivery tactics. It is usually determined by a sonographer—a specialist with years of training working on expensive stationary equipment—so high-quality obstetric ultrasound is often inaccessible in low-resource regions. The blind sweep protocol simplifies imaging: an operator without special training simply moves the sensor over the abdomen without performing manual measurements, while the model calculates the gestational age from the recording. The system used in the study was previously trained on 1,968 participants from the US, where data collection took place in North Carolina, and Zambia—using Sonosite, GE, and Butterfly Networks sensors. It was minimally fine-tuned for the new portable Clarius C3 sensor on only 120 participants from Chicago, which is approximately an order of magnitude less data than required for the initial training.

Why this matters for the industry

For the industry, this is the first prospective multicenter test of AI ultrasound transferability simultaneously to a new hardware platform and a new population without local fine-tuning. The value here is more methodological than product-oriented: the study sets a template for proving transferability—separate cohorts by center, a pre-specified non-inferiority criterion, and comparison with experts on the same data. The recipe for economical adaptation was also confirmed: fine-tuning for a new sensor took 120 scans compared to 1,968 in the initial training. The combination of a simple data capture protocol with a model transferable between sensors proves viable on cheap portable probes and minimally trained operators, providing a ready-made architectural hint for teams building medical AI: such an external validation design should be adopted as an internal standard of proof and used to benchmark their own models. A likely, but not guaranteed, consequence will be pressure on portable probe vendors to embed AI inference into their own applications.

Why this matters for users

The practical significance for readers is that expert-level gestational age assessment can be provided not by a sonographer with years of training, but by a paramedic or nurse after minimal training, working with a portable sensor. For patients in regions with a shortage of specialists, this is a chance for correctly determined gestational age even where there is no staff sonographer or stationary machine, and with it—proper planning of pregnancy management and earlier detection of complications. The scale of the problem the technology targets is significant: according to WHO data, lack of access to diagnostics is associated with more than 250,000 preventable maternal deaths annually. If the transferability of such a solution is confirmed on new sensors and in new populations, the 'non-specialist, portable sensor, and AI' scenario could become a standard component of prenatal screening in low-resource regions.

What is still unknown / limitations

Several points need clarification. First, exactly one pair of 'new hardware and new population' was tested: one Clarius C3 sensor and one fine-tuning on 120 participants, so a general property of AI model transferability between hardware platforms cannot be inferred from the result. Second, the mean error of 4.18 days does not show the tails of the distribution, and in prenatal practice, it is precisely the gross errors in gestational age estimation that are clinically dangerous; their frequency is not disclosed in the sources. Third, the results were obtained in research conditions at two centers, Northwestern Medicine and Jacaranda Health; outside these protocols, the system does not operate, and no product or API based on it exists. Finally, expectations of a wave of reproductions on other sensors and regulatory discussions on transferring medical AI models between devices are predictions, not confirmed facts.

Sources

Author

Look at AI, editorial team