What AI-assisted ultrasound can measure, and what it has not proved

A simpler way to collect the video

Google's October 6 interview about prenatal ultrasound describes an approach designed for places where trained sonographers are scarce. Healthcare workers move a portable probe over the abdomen in a predefined pattern, known as a blind sweep. AI then interprets the collected video instead of requiring the operator to capture each precise measurement.

The researchers discuss estimating gestational age, meaning how far along a pregnancy is, and identifying fetal presentation, or position. Those are defined tasks. They are not the same as checking everything a clinician may look for during a prenatal examination.

This is a closer look at the evidence behind the interview, not a newly launched diagnostic service.

The main comparison involved 385 people

The JAMA Network Open study, published on July 9, drew from a larger group of 2,043 participants in Chicago and Nairobi. Its primary age-estimation analysis covered 385 people at 16 to 36 weeks of pregnancy. The model's mean absolute error was 4.2 days, compared with 4.5 days for the clinical standard. The authors found it met their one-day noninferiority margin.

That means it passed the study's chosen test for not being meaningfully worse. It is not a finding that AI replaces sonographers. An average error also does not describe the error in every individual scan.

Offline processing does not remove every access barrier

Google says the processing runs on the device without Wi-Fi. Its earlier technical account, from June 2023, explains how mobile optimisation lets video models run locally and provide feedback when a sweep is poor. A portable system can reduce dependence on a live internet connection.

It still needs suitable hardware, power for the device, trained operators and a route into clinical care. Google describes a research application and a possible way to broaden access. Neither that description nor a laboratory framework establishes that a service is available in every clinic.

Accuracy is only one part of clinical evidence

The paper reports differences in operator training, limited testing across hardware and uncertainty in some subgroups and later pregnancies. Those gaps matter when a system moves beyond the settings in which it was evaluated.

The measured result is pregnancy-age estimation, not a demonstrated reduction in deaths or complications. Position estimates discussed in Google's interview should not be mistaken for a claim that the July study validated every prenatal diagnostic task.

A useful deployment would need more than a good average score: reliable scan collection, careful handling of uncertain results and evidence that the information helps people receive appropriate care. The studies are a reason to investigate that possibility, not to declare the clinical problem solved.