Ultrasound AI highlights exploratory early-birth predictions from Chile study
The company’s November 18 account highlights results for premature deliveries. The underlying study found a promising risk signal, but no patient’s care used the model.
Loading page…
The company’s November 18 account highlights results for premature deliveries. The underlying study found a promising risk signal, but no patient’s care used the model.
Listen to this story
Results from a four-hospital Chilean study sharpen the distinction between estimating a delivery date and identifying who may deliver early: after accounting for local birth timing, Ultrasound AI’s date-prediction advantage over conventional estimates shrank to less than one day. A separate risk score showed stronger discrimination in a prespecified analysis (AUC 0.86 versus 0.52 for gestational age alone), but was miscalibrated and tested retrospectively, with no AI-guided care. The findings support further validation, not clinical deployment; the product is not authorized for clinical use or sale in Chile, and the company says its US version is not intended to predict preterm birth.
Researchers applied the unmodified image-only model to 64,235 frames from 3,788 examinations across 2,550 pregnancies at four public hospitals; 34% involved migrant mothers.
Outside the restricted prespecified analysis, the risk score exceeded a baseline using coded diagnoses and same-examination ultrasound measurements by 0.09, but estimated risks did not reliably match outcomes.
The research was published in npj Digital Medicine on November 5; News Medical published the company-supplied account on November 18.
Ultrasound AI highlighted a potential early-birth warning from routine pregnancy scans in a November 18 company-supplied article published by News Medical. For pregnancies ending in premature delivery, it said predicted birth dates fell within two weeks of actual delivery 71% of the time. Those results came from a retrospective study, not a test of AI-guided care.
The underlying research appeared November 5 in npj Digital Medicine. Led by Rodrigo Alliende-Ferrada, researchers applied the commercial, image-only model without modification to 64,235 frames from 3,788 examinations. The study covered 2,550 pregnancies at four Chilean public hospitals; 34% involved migrant mothers.
The model initially appeared substantially better at predicting delivery dates. But the authors found that most of its advantage reflected shorter pregnancies in the local population. After adjusting comparison dates for that local timing, the gain fell below one day.
Average size of the model’s delivery-date error, regardless of direction.
The initial comparison, before adjustment for local birth timing.
A conventional estimate assigning every pregnancy the same gestational age at birth cannot rank who will deliver early. It can still produce a small average date error when most births occur near term. The researchers therefore examined a separate image-only risk score.
In a prespecified analysis restricted to a defined point in pregnancy, that score achieved an area under the curve of 0.86, versus 0.52 for gestational age alone. This measures how well it distinguishes early births from other births—not the percentage of individual predictions it gets right.
Outside that restriction, it exceeded a baseline combining coded diagnoses and same-examination ultrasound measurements by 0.09. Yet it was miscalibrated: its estimated risks did not reliably match observed outcomes. The authors called these classification analyses exploratory; no patient was managed using the model, and prospective testing remains necessary.
The paper says the product lacks authorization for clinical use or sale in Chile. The company’s November 18 account also states that, in the United States, Delivery Date AI is not intended to predict preterm birth.
Loading discussion...
Join the conversation
Explain what would make that warning helpful—or unsettling.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.