Google study finds AI ultrasound non-inferior across Chicago and Nairobi

A Google-backed prospective study found that an AI system met its prespecified non-inferiority standard for estimating gestational age from blind-sweep ultrasound videos collected by novice operators in Chicago and Nairobi. The primary evaluation covered 385 participants from 16 through 36 weeks of gestation.

The study, published in JAMA Network Open, reported a pooled mean absolute error of 4.2 days for the adapted AI model and 4.5 days for the clinical standard. That standard used fetal biometry performed by expert sonographers. With a one-day margin, the one-sided test returned P<.001. The result establishes non-inferiority under the study design, not superiority.

The evaluation included 192 participants at Northwestern Medicine in Chicago and 193 at Jacaranda Health in Nairobi. Five operators with no previous ultrasound experience captured B-mode sweeps with Clarius probes. A research app on tablets processed the videos and generated predictions on-device in real time. Operators did not see those predictions during the study.

Training differed between sites. The two Chicago operators received a handout, verbal instructions and a brief demonstration. The three Nairobi operators received six hours of hands-on training, which also covered study procedures and shadowing expert sonographers.

Investigators adapted a model originally trained on data from North Carolina and Zambia and on images from other ultrasound devices. They fine-tuned it with 180 examinations from 120 Chicago participants to support Clarius C3 video. Because no Nairobi data were used for fine-tuning, the Nairobi evaluation was an external test of the Chicago-adapted model rather than a locally tuned system. In Chicago, site-level mean absolute error was 4.1 days for the model and 4.6 days for the clinical standard. Both figures rounded to 4.3 days in Nairobi.

The model rejected 4.9% of sweeps overall, including 7.9% in Chicago and 1.8% in Nairobi. The authors said different training protocols could have contributed to that gap, but the study did not isolate the cause. Beyond 36 weeks, exploratory performance was weaker in Nairobi: among 54 cases, mean absolute error was 12.0 days for the model and 5.0 days for the clinical standard.

Google Research described the work on October 21. Google funded it through grants to Jacaranda Health and Northwestern University, with partial Gates Foundation funding through grants to Northwestern. Google employees participated in study design, analysis, interpretation, manuscript preparation and the submission decision. Several authors disclosed Google employment and stock ownership.

The study measured gestational-age estimation error. It did not test whether the system expanded access to prenatal care, improved safety, gained clinical adoption or changed maternal or neonatal outcomes.