The World Health Organization estimates that more than 250,000 preventable pregnancy-related deaths occur worldwide each year. Prenatal ultrasound can improve outcomes and reduce fetal and maternal mortality, yet two-thirds of people globally lack access to diagnostic medical imaging such as ultrasound. Current ultrasound-based maternal care depends heavily on specialized sonographer training, leaving millions in low-resource areas without any ultrasound during pregnancy.
To address this gap, researchers developed machine-learning systems that interpret standardized "blind sweep" ultrasound scans. In this scanning protocol, frontline health workers without formal sonography training glide a low-cost, portable ultrasound probe across the abdomen in a predefined pattern; the recorded image sweeps are then analyzed by an AI model.
Because many regions face shortages of trained experts and because sonographer training is costly and multi-year, enabling minimally trained health workers to perform assessments and relying on AI interpretation could materially expand access to prenatal screening and improve maternal and fetal care.
The study
The paper "Generalization of AI-Based Gestational Age Assessment Using Blind Sweep Ultrasonography" was published in JAMA Network Open. It evaluated how well an AI model for gestational age assessment generalizes across different clinical environments and hardware when used prospectively in real clinics on mobile devices by novice operators.
The original model was only minimally tuned on one target population and then prospectively tested across two sites:
- Northwestern Medicine: an urban, high-resource academic medical center in Chicago, IL, United States.
- Jacaranda Health: an urban community clinic network in Nairobi, Kenya.
The model had been trained on ultrasound data collected in North Carolina and Zambia using Sonosite, GE, and Butterfly Networks commercial probes. For the prospective tests, the research system was paired with a Clarius C3 handheld transabdominal probe. The study sites differed in care patterns, such as timing of first prenatal visits, the number of routine ultrasounds, and standard methods for dating gestational age.
To assess efficient tuning, researchers used a small pilot dataset of 120 participants from Chicago—about 6% of the original training set—to adapt the model to the new probe hardware and to extend the gestational age prediction range beyond 36 weeks. No data from the Nairobi site were used for fine-tuning, so the Kenyan validation served as an independent external test. The study also included multiple-gestation pregnancies to evaluate performance in more complex measurement scenarios.
Key findings: comparable to expert standard
The study compared the AI’s gestational age estimates from blind-sweep ultrasounds to formula-based dating from the clinical standard performed by expert sonographers. The expert approach requires localization of multiple fetal cross-sections and precise biometry measurements—skills that require years of training and are operator-dependent.
In the primary cohort of 385 participants between 16 and 36 weeks' gestation, the diagnostic study reported:
- Clinically non-inferior accuracy: the AI model achieved a mean absolute error (MAE) of 4.18 days for gestational age estimation, compared with an MAE of 4.48 days for the expert sonographer standard, demonstrating non-inferiority.
- Cross-setting generalization: after adaptation, the Chicago site yielded an MAE of 4.10 days; in Nairobi, with no local optimization, the model achieved an MAE of 4.27 days.
Exploratory analyses indicated robustness in challenging situations as well: for fetuses with weights above the 90th percentile, the AI outperformed traditional biometry formulas, suggesting resilience to variations in fetal size.
Operational lessons from prospective deployment
Beyond demonstrating potential to expand ultrasound access, the multi-center prospective deployment yielded practical operational insights. Deploying the system in routine clinical workflows revealed needs such as swappable batteries rather than plug-in docking stations to keep devices charged between consecutive patients.
The authors also noted that deploying such AI tools will require building deep trust with clinicians and patients. Practical safeguards discussed include model-produced confidence estimates to reject poor-quality sweeps and communicate uncertainty to the clinical user.
Implications and next steps
The findings show that AI can analyze standardized blind-sweep image sequences acquired by operators without specialized sonography training, producing gestational age estimates comparable to current expert-operated standards and cart-based ultrasound systems. Demonstrating technical generalizability across diverse settings, hardware, and populations marks an important step toward tools that could broaden access to prenatal ultrasound.
By enabling minimally trained community health workers to perform reliable ultrasound assessments, this approach could expand prenatal coverage for millions in under-resourced areas and contribute to more resilient, accessible healthcare systems that improve maternal and fetal outcomes.
Acknowledgements
The study’s co-authors include researchers from Google Research, Jacaranda Health, and Northwestern University Feinberg School of Medicine: Chace Lee, Justin Krogue, Anneka Wickramanayake, Stacey Caron, Priscah Cheruiyot, Amber Watters, Alicia Martín, Tiya Tiyasirichokchai, Akib Uddin, Yun Liu, Patricia Strachan, Sunny Jansen, Alyssa Santi, Preeti Singh, Catherine Arguelles, Kenneth Hudson, David Melnick, Mahesh Vaidyanathan, Mozziyar Etemadi, Yossi Matias, Avinatan Hassidim, Greg Corrado, Shruthi Prabhakara, Daniel Golden, Ryan G. Gomes, and Shravya Shetty. The authors also thank Pi-Chuan Chang and Dale Webster for reviewing the work.
The reported results are drawn from the prospective, multi-center diagnostic study detailed in the JAMA Network Open publication.
