External validation of an image-only ultrasound model for delivery-date prediction in four Chilean public hospitals

Abstract

Abstract Accurate delivery-date estimation underpins obstetric care, yet conventional dating performs worst where it is most needed: late-presenting pregnancies without reliable early ultrasound or menstrual history, common in public-sector and migrant populations. We report an independent, multicentre external validation of a prospectively deployed, image-only ultrasound artificial-intelligence (AI) system that predicts days to delivery from standard ultrasound frames alone—without gestational age, biometry, demographics, annotations, or any dating input. Across 3,788 singleton studies from 2,550 pregnancies at four Chilean public hospitals, the AI predicted the delivery date with a mean absolute error (MAE) of 6.9 days (95% confidence interval [CI] 6.7–7.0). Critically, its accuracy was undiminished where conventional dating is structurally weakest: in pregnancies lacking reliable first-trimester dating the AI achieved an MAE of 6.4 days, whereas estimation from the established due date deteriorated to 13.2 days. The model’s high-coverage prediction intervals were well calibrated (96.5% empirical coverage of the approximately 94% interval), supporting a calibrated delivery-date range rather than a single date. Head-to-head within the same studies, the image-only estimate was more accurate than the due date as used in practice in 69% of pregnancies; prespecified sensitivity analyses showed that part of this margin reflects correction of a systematic, population-specific bias in the 280-day convention, with a smaller residual advantage in dispersion. This dating-independent, calibrated estimate could narrow a persistent equity gap for the patients whom conventional dating serves least well.