Novice operators using a handheld probe obtained pregnancy-age estimates with average error comparable to conventional ultrasound dating in a 385-participant primary evaluation. The AI interpreted clips from standardised abdominal sweeps, without requiring the operator to locate and measure individual anatomical structures.[1][2]
Researchers from Google, Northwestern University and Jacaranda Health evaluated the method in a prospective diagnostic comparison in Chicago and Nairobi. It met the prespecified non-inferiority criterion against the conventional clinical dating approach.[1]
An unusually clear comparison
The primary analysis involved 385 pregnant participants: 192 in Chicago and 193 in Nairobi. The team used five novice operators, two at the Chicago site and three in Nairobi. Chicago operators received a handout, verbal instruction and demonstration; those in Nairobi completed six hours of formal training. The clinical question was narrow and measurable: how close was each estimate to the pregnancy age established by the study’s reference method?[1]
Across the combined primary evaluation, the AI’s mean absolute error was 4.2 days (95% confidence interval 3.8–4.6), compared with 4.5 days (4.1–4.9) for the clinical standard. The prespecified one-day non-inferiority margin was met. These are mean absolute errors across the evaluation group; an individual estimate can differ by more or less than those averages.[1]
What the clinical study found
| Measure | Value | Context |
|---|---|---|
| Primary evaluation participants | 385 (Chicago 192; Nairobi 193) | |
| Novice operators | 5 across the two centres | |
| Formal operator training in Nairobi | 6 hours | |
| AI mean absolute dating error | 4.2 days overall | |
| Clinical-standard mean absolute error | 4.5 days overall | |
| What was compared | Accuracy of pregnancy-age dating, not delivery outcomes |
A model leaves home
A clinical AI system can look brilliant on familiar hospital data and lose its bearings when used with different patients, devices or routines. That is why the Nairobi cohort matters. The model was developed using other datasets, and the paper examines the system in a setting distinct from the US site, with different operators and equipment. The Nairobi results were not simply a demonstration in the original laboratory. There was still a controlled research protocol, however, and the study describes model fine-tuning associated with the Chicago setting.[1]
The researchers also report rejected sweeps—roughly 1.8% at Nairobi and 7.9% at Chicago—rather than treating every recording as usable. This is a practical detail with considerable implications: a field system needs to recognise when an image is inadequate and ask for another acquisition. A method that appears accurate only after silently discarding its hardest examinations may not deliver the access its designers hope for. The reported rejection behaviour therefore belongs in the story alongside the successful predictions.[1]
What the access result establishes
The demonstrated capability is gestational-age estimation from novice-acquired sweeps in two clinical settings. It reduces the acquisition skill required for that measurement. Google’s October 21 account describes this as a route toward wider maternal-care access; the study did not establish improved birth outcomes.[2][1]
The next operational question is how to deliver the measurement consistently: obtain an adequate sweep, recognise when it needs repeating and integrate the result into local antenatal care. Comparable dating error is a concrete starting point for that work.[1]
Sources & further reading
- Generalization of AI-Based Gestational Age Assessment Using Blind Sweep Ultrasonography Willis and collaborators / JAMA Network Open · 9 Jul 2026 · Accessed 24 Oct 2026
- Empowering frontline maternal care with AI-assisted ultrasound Google Research · 21 Oct 2026 · Accessed 24 Oct 2026