Artificial intelligence may improve accuracy of gestational age estimation

Artificial intelligence models trained to analyze ultrasonograms may help sonographers more accurately estimate gestational age, according to data published in JAMA Network Open.

“Fetal ultrasonography is the cornerstone of prenatal imaging and provides crucial information to guide maternal-fetal care, such as estimated gestational age and evaluation for fetal growth disorders,” Chace Lee, MS, of Google Health in Palo Alto, California, and colleagues wrote. “Currently, the clinical standard for estimating gestational age and diagnosing fetal growth disorders is determined through manual acquisition of fetal biometric measurements, such as biparietal diameter, head circumference, abdominal circumference, femur length or crown-rump length.”

Since estimation of gestational age is reliant on sonographers’ skill and experience, researchers have worked to develop AI systems to assist sonographers in their estimations, according to study background.

Methods

For the present study, Lee and colleagues developed three AI models to determine gestational age. The first model predicted gestational age in days based on fetal ultrasonography images captured during biometry measurements, the second analyzed 5- to 10-second fly-to videos captured immediately before image capture and the third — termed the ensemble model — analyzed both images and fly-to videos. All models used the Hadlock regression formula for estimations.

The models were trained and evaluated on prospective data from the Fetal Age Machine Learning Initiative (FAMLI) study, which enrolled pregnant people who received antenatal care in Chapel Hill, North Carolina, or Lusaka, Zambia. FAMLI sonographers manually evaluated gestational age using ultrasonography devices with the built-in AI turned off. The researchers randomly assigned participants to the training data set (60%), the tuning data set (20%) and the testing data set (20%).

Results

The testing data set included 404 participants (mean age, 28.8 years) who collectively attended 677 study visits.

The image-based model more accurately estimated gestational age compared with standard biometry (mean difference in mean absolute error (MAE), –1.13 days; 95% CI, –1.5 to –0.7), as did the video-based model (mean difference in MAE, –1.48 days; 95% CI, –1.9 to –1.1) and the ensemble model (mean difference in MAE, –1.51 days; 95% CI, –1.9 to –1.1).

Subgroup analyses indicated that the AI models generalized well across the second and third trimesters, devices and countries.

“We found that in the third trimester, our model’s accuracy advantage relative to the clinical standard fetal biometry increased,” Lee and colleagues wrote. “This is particularly important because accurate gestational age estimation in the third trimester is essential for managing complications and making appropriate clinical decisions regarding timing of delivery.”

Additionally, the ensemble model had lower MAE compared with the National Institute of Child Health and Human Development (mean difference, –1.23 days; 95% CI, –1.6 to –0.8) and the Intergrowth-21st (mean difference, –2.69 days; 95% CI, –3.3 to –2.1) regression formulas.

“Sonographers are in high demand and often have workplace or overuse injuries due to current scanning requirements,” Lee and colleagues wrote. “Additional studies are needed to investigate whether an AI adjunct can reduce scanning time, assist sonographers and minimize workplace injury.”