AI-based ultrasound model shows high accuracy in predicting delivery timing

AI-based ultrasound model shows high accuracy in predicting delivery timing | Image Credit: © Gorodenkoff - stock.adobe.com.

A new artificial intelligence (AI) platform developed by Ultrasound AI has demonstrated high accuracy in predicting delivery timing from standard obstetric ultrasound images, according to findings from the Perinatal Artificial Intelligence in Ultrasound (PAIR) study published in The Journal of Maternal-Fetal & Neonatal Medicine. The study, conducted with researchers at the University of Kentucky, suggests the technology could improve risk assessment for preterm birth (PTB) and enhance maternal-fetal care, particularly in resource-limited settings.^1,2

“This is a major milestone for the field of maternal-fetal medicine and for Ultrasound AI,” said Robert Bunn, founder and president of Ultrasound AI. “Our AI’s ability to accurately predict delivery timing—and learn and improve over time—has profound implications for both clinical practice and public health, especially in settings where early risk identification is critical and access to specialist care is limited.”

Study design and methods

The retrospective cohort study analyzed de-identified ultrasound images from 5,714 pregnant patients who delivered at the University of Kentucky between 2017 and 2021. A total of 19,940 unique ultrasound exams and 877,141 images were used for initial AI training. Of these, 79% were allocated to a training set and 21% to an independent validation set.

The AI used deep learning algorithms, including convolutional neural networks and transformers, to process still ultrasound images without input from clinical measurements, maternal history, or operator annotations. Predictions focused on the number of days until delivery and whether the birth would be preterm (<37 weeks’ gestation). The model underwent multiple retraining iterations, incorporating more than 2 million ultrasound images by the final version (V4).

Key results

In the initial model (V1), the AI achieved an R² of 0.90 for term births and 0.85 for all births combined when predicting days to delivery. For spontaneous PTB, the R² was 0.48. Retraining improved performance substantially: in V4, R² reached 0.95 for term births, 0.92 for all births, and 0.72 for spontaneous PTB.

The model’s PTB prediction sensitivity was initially 39%, with a specificity of 93% and an area under the receiver operating characteristic curve (AUC) of 0.757. After retraining, sensitivity increased to 40% and specificity to 95%, with AUC rising to 0.825. For births occurring within 30 days, AUC improved from 0.842 to 0.900.

Performance was consistent across trimesters, with mean absolute error (MAE) ranging from approximately 12 to 15 days. The AI maintained accuracy across patient demographics, including variations in age, body mass index, and race or ethnicity.

Clinical implications

Preterm birth remains the leading cause of neonatal mortality worldwide, and accurate prediction is a persistent challenge. Traditional PTB risk assessment methods rely on clinical history, biomarker testing, and cervical length measurements, which can be limited by interobserver variability and access to specialized care.

Ultrasound AI’s platform offers a noninvasive, operator-independent tool that can be integrated into standard imaging workflows without additional testing. “AI is reaching into the womb and helping us forecast the timing of birth, which we believe will lead to better prediction to help mothers across the world,” said John M. O’Brien, MD, division director of maternal-fetal medicine at the University of Kentucky.

The authors noted that continuous learning capabilities and the ability to process large imaging datasets enable the model to improve over time. The system could potentially be enhanced by integrating serial ultrasound exams, biochemical biomarkers, and Doppler or fetal movement data.

Conclusion

The PAIR study demonstrates that AI can predict delivery timing with high accuracy using only ultrasound images and can improve the prediction of preterm birth with iterative retraining. Further validation in diverse populations and integration into clinical workflows could advance personalized obstetric care, potentially improving maternal and neonatal outcomes.

References:

1. Hiltzik Strategies. Ultrasound AI publishes landmark study demonstrating breakthrough in predicting delivery timing using AI and ultrasound images. August 13, 2025. Accessed August 14, 2025. https://www.eurekalert.org/news-releases/1094602
2. Patel N, O’Brien J, Bunn R, Schanbacher B, Bauer J, Lam GK. Perinatal artificial intelligence in ultrasound (PAIR) study: predicting delivery timing. The Journal of Maternal-Fetal & Neonatal Medicine. 2025;38(1). doi:https://doi.org/10.1080/14767058.2025.2532099