Deep learning to estimate gestational age from blind ultrasound sweeps of the gravid abdomen

Background: Ultrasound is indispensable to gestational age estimation, and thus to quality obstetric care, yet high equipment cost and need for trained sonographers limit its use in low-resource settings. Methods: From September 2018 through June 2021, we recruited 4,695 pregnant volunteers in North Carolina and Zambia and obtained blind ultrasound sweeps (cineloops) of the gravid abdomen alongside standard fetal biometry. We trained a neural network to estimate gestational age from the sweeps and, in three test sets, assessed performance of the model and biometry against previously established gestational age. Results: In our main test set, model mean absolute error (MAE) was 3.9 days (standard error [SE] 0.12) vs. 4.7 days (SE 0.15) for biometry (difference -0.8 days; 95% CI -1.1, -0.5; p<0.001). Results were similar in North Carolina (difference -0.6 days, 95% CI -0.9, -0.2) and Zambia (-1.0 days, 95% CI -1.5, -0.5). Findings were supported in the test set of women who conceived by in vitro fertilization (model MAE 2.8 days [SE 0.28] vs. 3.6 days [SE 0.53] for biometry; difference -0.8 days, 95% CI -1.7, 0.2), and in the set of women from whom sweeps were collected by untrained users with low-cost, battery-powered devices (model MAE 4.9 days [SE 0.29] vs. 5.4 days [SE 0.28] for biometry; difference -0.6, 95% CI -1.3, 0.1). Conclusions: Our model estimated gestational age more accurately from blindly obtained ultrasound sweeps than did trained sonographers performing fetal biometry. These results presage a future where all pregnant people - not just those in rich countries - can access the diagnostic benefits of sonography.