Complications of preterm birth are the leading cause of death among children under five worldwide, according to the World Health Organization. Every pregnancy gets a due date, but a due date assumes every baby arrives at 40 weeks. Blood tests can flag a higher risk of preterm birth, but they report a risk, not a date.
A new peer-reviewed study tested whether routine ultrasound images alone can do more. Led by Dr. Rodrigo Alliende-Ferrada, the investigators ran Ultrasound AI's technology, unchanged, on past scans from 2,550 pregnancies at four public hospitals in Chile. About a third of the mothers were migrants.
What they found:
- For babies who came early, the AI's predicted delivery date was within two weeks 71% of the time, and off by about 11 days on average.
- Take two pregnancies at the routine 23-week scan, one that went on to deliver early and one that didn't. Using only the images, the AI picked the one headed for early birth 86% of the time. A coin flip gets 50%.
- The warning showed up both for natural early labor and for early deliveries that doctors scheduled for medical reasons.
- Across all pregnancies, the predicted date was within two weeks of the actual birth more than 9 times in 10. It held up for women who didn't have an early scan, often the mothers with the least access to care, and accuracy was similar for Chilean and migrant mothers.
This was a look back at past pregnancies, and no patient's care was based on the AI. The investigators describe the early-birth results as exploratory; the next step is a forward-looking study. In the United States, Delivery Date AI is not intended to predict preterm birth.
Ten years ago I started working on a simple question: does an ultrasound image hold more than the few measurements we take from it? This study, led by doctors at four public hospitals in Chile, suggests it does. Knowing when a baby may come early could give mothers and their care teams time to prepare."
Robert Bunn, founder and CEO of Ultrasound AI
Dr. Alliende-Ferrada serves as a strategic advisor to Ultrasound AI. Ultrasound AI gave the investigators access to its web platform to upload de-identified images.
Source:
Journal reference:
Alliende-Ferrada, R., et al. (2026). External validation of an image-only ultrasound model for delivery-date prediction in four Chilean public hospitals. npj Digital Medicine. DOI: 10.1038/s41746-026-03453-0. https://www.nature.com/articles/s41746-026-03453-0