At Northwestern Medicine in Chicago and Jacaranda Health in Nairobi, people who had never done an ultrasound held a handheld ultrasound device against pregnant women's bellies and swept it in set directions. The device is a probe that you move across the abdomen. The screen showed only images of the inside of the belly, and an AI running on a tablet estimated gestational age from those images alone. The AI's error was similar to or smaller than the error of expert sonographers who scanned the same women.
Expert sonographers scanned the same women separately, and both sets of estimates were checked against gestational age confirmed by an early pregnancy scan
The study is a prospective diagnostic study published in JAMA Network Open in July 2026. Researchers enrolled 2,043 people across the two cities and used 385 of them, at 16 to 36 weeks of pregnancy, for the primary analysis (192 in Chicago, 193 in Nairobi). Each woman had two exams. In one, a person with no ultrasound experience swept her belly with the handheld device and the AI estimated gestational age. The other was a routine clinical ultrasound within a week of the first. In the routine exam, an expert sonographer found fetal planes, took measurements, and calculated gestational age with the Hadlock formula. Both estimates were compared against the gestational age from an early pregnancy dating scan, which served as the reference.
Five people with no ultrasound experience held the device. The two in Chicago had no formal training, only written instructions, a verbal explanation, and a short demonstration. The three in Nairobi had six hours of hands-on training. The exam went in this order.
- Measure fundal height
Based on fundal height measured over the belly, set the imaging depth to 11 cm, 13 cm, or 15 cm. - Sweep in set directions
Sweep the Clarius C3 handheld probe up to six times: three times vertically and three times horizontally. This method is called a blind sweep. - AI reads the images on the device
The AI in the tablet app estimates gestational age directly from the video. The operator sees the live images but not the AI's estimate.
Google LLC funded the work at Northwestern University and Jacaranda Health, with part of the funding from the Gates Foundation. Google-affiliated authors took part in design, analysis, interpretation, and writing, while the two sites handled data collection. Fourteen of the authors are Google employees and shareholders, and three hold Google patents on the approach that combines blind sweeps with AI. Every number below comes from a study designed and analyzed by the company that built the technology.
Average error in gestational age: 4.2 days for the AI, 4.5 days for expert sonographers
Error here means the average number of days between the estimated and reference gestational age. The AI came in at 4.2 days (95% CI, 3.8–4.6 days) and the expert standard procedure at 4.5 days (4.1–4.9 days). Before the study began, the researchers set a rule: “the AI is noninferior if its error is no more than 1 day larger than the standard procedure's.” The AI met that rule.
4.2 days is smaller than 4.5, but the primary test asked only whether the AI was worse by more than 1 day. The primary analysis did not test whether the AI was more accurate, and the two confidence intervals overlap. The researchers also checked whether errors leaned one way. The standard procedure tended to underestimate gestational age by an average of 0.7 days (−0.7 days), and this bias was statistically significant. The AI underestimated by 0.2 days on average (−0.2 days), which was not significant.
| Group | Mean error, AI (blind sweep) | Mean error, expert sonographer standard procedure |
|---|---|---|
| Overall (385) | 4.2 days (3.8–4.6) | 4.5 days (4.1–4.9) |
| Chicago (192, model tuned for the new device) | 4.1 days (3.5–4.7) | 4.6 days (4.1–5.3) |
| Nairobi (193, no local tuning) | 4.3 days (3.8–4.8) | 4.3 days (3.8–4.9) |
The Google Research blog reported the same result as 4.18 days versus 4.48 days and described it this way. The blog was published by the company that developed the technology.
“comparable to the current standard of care from expert operators and more expensive traditional, cart-based ultrasound machines”
The AI was never trained on Nairobi patients, yet in Nairobi its average error matched the experts' exactly: 4.3 days
The AI was originally trained on data from pregnant women in Chapel Hill, North Carolina, and Lusaka, Zambia (1,968 people, according to the blog). Those scans used standard ultrasound machines and the ButterflyIQ. Before this study, the only tuning used data from 120 participants in a Chicago pilot study. The goal was to adapt the model to images from the newly used Clarius C3 probe and to add some images from beyond 36 weeks. The blog says these 120 people were about 6% of the original training data.
Nairobi data was never used for training or tuning. The blog calls this “zero local optimization”. Even so, among the 193 women in Nairobi, the AI's mean error was 4.3 days, and the expert standard procedure on the same women was also 4.3 days. At a hospital and with patients it had never seen in training, the AI matched the experts' error.
The AI did not need Nairobi data to match the experts, but there are signs that a new device did call for adjustment. When the original model, before tuning for the Clarius C3, was run on 53 Chicago participants, the AI's error was 5.1 days, larger than the standard procedure's 4.2 days. After tuning, the error across 192 Chicago participants was 4.1 days. These two figures come from subgroup analyses with different participants, though, so they are weak evidence that tuning for the device reduced error.
The paper lays out limitations alongside these results. Nairobi had no local tuning, so there may be room to improve performance there. Clarius was the only new device manufacturer tested. In the late-term range past 36 weeks, outside the primary analysis, the AI's error in Nairobi was 12.0 days, more than twice the standard procedure's 5.0 days (54 women). The researchers gave two possible reasons. One is that the expert sonographers already knew the gestational age from the early dating scan used as the reference when they scanned. If they scanned with that number in mind, the experts' values may have come out closer to the reference. The other is that the model may need improvement in the late-term range.
The AI doesn't measure the baby the way experts do. It reads gestational age straight from the video, so the person only has to sweep as instructed
In the standard procedure, an expert sonographer watches the screen to find fetal planes, measures things like head and abdominal circumference and femur length, and plugs them into the Hadlock formula to calculate gestational age. Finding the right planes and measuring accurately takes skill.
The AI in this study skips the measurements and estimates gestational age in days directly from the sweep video. It scores confidence by how much its predictions vary across each video, rejects low-quality sweeps, and combines predictions from the remaining videos, giving more weight to those with higher confidence. Operators swept without seeing the result, so in this study the person's only job was to sweep at the set depth and in the set directions.
Because the methods differ, they also go wrong in different cases. In an exploratory analysis of large babies, with estimated fetal weight above the 90th percentile, the standard procedure's error rose to 7.2 days, while the AI's was 4.4 days (P=.004). The sample was small, and the researchers treated it as an exploratory finding. Instead of skill in measuring, this approach depends on the quality of the sweep video. A separate research group simulated mistakes such as sweeping in the reverse direction, flipping the probe, or stopping partway, and reported that these deviations affect how blind sweep AI classifies fetal presentation and placental location. They proposed detecting quality problems and rescanning. That study did not test gestational age estimation or Google's model.
Other tools use the same approach. Butterfly Network's blind sweep gestational age tool, which uses a model from a University of North Carolina research team, received FDA clearance in March 2026. According to the article, it was validated for 16 to 37 weeks and produces an estimate in under two minutes.
What the Nairobi operators learned in six hours of hands-on training wasn't how to read scans. It was how to sweep evenly, at a quality the AI could read
The six hours of hands-on training for the three Nairobi operators consisted of watching demonstration sweeps, practicing on phantoms and volunteers, and shadowing expert sonographers. The two in Chicago received only written instructions, a verbal explanation, and a short demonstration. The public description does not say who ran the training.
The share of sweeps the AI rejected as low quality was more than four times higher in Chicago than in Nairobi, and the researchers attributed the gap to the six hours of hands-on training. Gestational age error, on the other hand, was similar: 4.1 days in Chicago and 4.3 days in Nairobi. The two cities differed in model tuning as well as training, so the effect of training alone cannot be separated out from the error figures. Confidence intervals for individual operators were also wide, and the researchers wrote that the results need to be confirmed with a more diverse group of operators.
The problems operators ran into in the field had more to do with the hardware than with reading scans. Seeing patients back to back left no time to recharge, so operators wanted swappable batteries rather than a charging dock that needs the device plugged in for a long time. The researchers also listed unstable power, device durability, retesting hardware and software after updates, and earning the trust of clinicians and patients as challenges.
Until now, knowing gestational age took an expert who could find fetal planes and measure them. In this study, people with no ultrasound experience swept the belly as instructed, and the AI read the video and dated the pregnancy as accurately as the experts. The study shows that even where there are no expert sonographers, the work can be split: people sweep evenly as instructed, and the AI does the reading.
Go deeper
Full paper in JAMA Network Open Covers the design, results by city, limitations, and conflict of interest disclosures. pmc.ncbi.nlm.nih.gov
Google Research blog post Diagrams the training, tuning, and validation process and includes field stories such as the battery problem. research.google
Study on blind sweep quality assessment Looks at how sweep mistakes affect AI classification and how rescanning restores performance. arxiv.org