Delivery date predicted by AI from ultrasound images alone: external validation in four Chilean public hospitals, and an advantage that mostly comes from a bias in the 280-day rule
A Chilean team tested, without retraining it, a commercial software that predicts the number of days left before delivery from ultrasound images alone, on 3,788 exams from 2,550 pregnancies in four public hospitals. The model is off by 6.9 days on average versus 11.6 days for the due date given by the 280-day rule, and by 6.4 versus 13.2 days when the pregnancy was not dated in the first trimester. But the authors themselves show that most of this gap disappears once the due date's bias in this population is simply corrected: the real gain is more modest than the headline figure, and it mainly concerns poorly dated pregnancies.
The context
The estimated due date (EDD) is calculated by adding 280 days, or 40 weeks, to the start of pregnancy. That starting point is estimated either from the last menstrual period or, better, from a first-trimester measurement of the embryo (crown-rump length, or CRL). Yet births rarely happen on the expected day, and many women have no early scan: pregnancy discovered late, late entry into care, migration. For them, dating is uncertain, and so is planning the delivery.
Ultrasound AI, a company based in Lexington, Kentucky, sells a model that directly predicts the time to birth from ultrasound images, with no measurements, gestational age or clinical data as input. The development study, PAIR (Patel et al., Journal of Maternal-Fetal & Neonatal Medicine, 2025), covered 5,714 patients from a single centre, the University of Kentucky. According to trade press, the product, called Delivery Date AI, received FDA authorisation through the De Novo pathway in 2026, as an adjunctive tool when dating by last period or biometry is unreliable. An independent evaluation outside the United States was missing.
The method
This is an external validation: a frozen model is applied to data it has never seen, from another population. The authors analysed after the fact the outputs of a system deployed from March 2021 to November 2025 in four public hospitals in the Santiago area: Barros Luco Trudeau (1,563 exams), Padre Hurtado (1,310), El Pino (527) and Dr. Luis Tisné Brousse (388). The report follows TRIPOD+AI, the good-practice checklist for publishing a prediction model.
The model takes the images of one exam (median of 14 images) and returns quantiles, that is, a prediction range: for example, "80% chance that birth occurs between day X and day Y". Its architecture is not described, as the software is proprietary. Of 4,515 exams linked to an outcome, the authors excluded 402 exams from twin pregnancies and 325 exams with insufficient data, keeping 3,788 exams. A third of pregnancies (33.4%) contributed more than one; confidence intervals are therefore computed with a pregnancy-level bootstrap (3,000 resamples), which avoids counting the same patient twice as if she were independent.
The population is diverse: 34.4% migrant women (including 14.6% Venezuelan and 8.4% Haitian), 16.6% suspected fetal growth restriction, 12.8% gestational diabetes. 12.5% of pregnancies had no reliable first-trimester dating. The comparator is the official clinical EDD. The main metric is the mean absolute error (MAE), the average gap in days between predicted and actual date, regardless of direction.
The results
In ML metrics: the model's MAE is 6.9 days (95% CI 6.7 to 7.0), versus 11.6 days (11.2 to 12.2) for the EDD, i.e. 4.8 days less (p < 0.001). The model beats the EDD in 69.2% of exams. In women without reliable dating, its MAE is 6.4 days versus 13.2. For preterm births (before 37 weeks, 262 exams), it is 10.8 days versus 37.5: unsurprising, since an EDD fixed at 40 weeks cannot anticipate an early birth. Error varies little between hospitals (6.0 to 7.7 days). The ranges are well calibrated: the nominal ~94% interval contains the true date in 96.5% of cases, with a mean width of 41.4 days.
The most important figure lies elsewhere. In this cohort, the standard EDD has a systematic bias of 10.8 days: the 280-day rule is offset from the actual birth dates of this population, and this is not an effect of inductions (bias of 10.8 days for spontaneous deliveries, 10.9 for induced ones). The authors therefore build an "oracle" EDD, corrected for this mean offset. Against it, the model's advantage drops to 0.7 days, a non-significant difference (p = 0.19). Restricted to spontaneous deliveries, a precision advantage remains: 6.3 days versus 8.2.
Clinical translation, per 1,000 exams: the model places the birth within one week of the actual date in 568 cases, versus 376 for the raw EDD; within two weeks in 924 cases versus 729; and it is off by more than four weeks in 3 cases versus 40. But a large part of this gain could be obtained by simply shifting all EDDs by the same amount, a local adjustment that requires no AI.
What is good
First, a true external validation: a frozen American model, applied in another country, in four public hospitals, over almost five years and 2,550 pregnancies, with a large migrant population. This is exactly the kind of study that is missing for devices authorised on the basis of a single centre.
Second, rare honesty: the authors publish the "oracle" analysis that reduces their own main result to a non-significant difference, and they say so in the abstract. They add sensitivity analyses (spontaneous deliveries, one exam per pregnancy, first and last exam) that all give MAEs close to 7 days.
Third, calibration of the ranges is reported hospital by hospital (96 to 99% coverage for the 94% interval), and the comparison within matched gestational-age windows shows the model works as well in poorly dated women (7.3 versus 7.1 days between 14 and 28 weeks). This is the population the tool is indicated for.
What is less good
First, a biased comparator. The raw EDD, offset by 10.8 days, is an easy opponent; the real competitor, fetal biometry measured during the same exam (Hadlock formulas), could not be tested because measurements were not archived. The authors acknowledge it. Without this comparison, we do not know whether the model adds anything over a well-interpreted standard ultrasound.
Second, a risk of shortcut learning that cannot be checked: a model that sees only images may mostly estimate fetal size, implicitly redoing biometry, rather than detecting maturation signs specific to late pregnancy. Because the software is closed (architecture, weights and training data unavailable), no explainability analysis was possible. Add a misleading metric: an R² of 0.97 looks spectacular, but it is inflated by the wide range of time-to-delivery values, from a few days to about six months depending on when the exam took place; the MAE and the percentages within 7 and 14 days are more telling.
Third, the limits of the setting: a retrospective study of a deployed system, with no measurement of its effect on decisions (inductions, transfers, maternity-ward planning); a single country and only the public sector; 727 exams excluded; no correction for multiple comparisons across subgroups. On independence, the authors declare no competing interests and state that the company neither funded nor steered the study, but the study had no dedicated funding, and in the preprint version we read, the links to the analysis code (Zenodo, GitHub) were still placeholders.
What it changes
For the research community: this is a useful reporting template, showing how to separate the part of a gain due to correcting a population bias from the part due to genuine added precision. The obvious next step is a head-to-head comparison with same-exam biometry, and a prospective study measuring clinical outcomes.
For clinicians: for a woman dated in the first trimester, the benefit is small, and a local recalibration of the EDD can capture most of it. For a woman who presents late without reliable dating, a calibrated range with about one week of mean error can help plan care, as a complement to clinical judgement and not a substitute. The 94% range remains wide (more than 40 days).
For patients and the public: AI does not "guess" the day of birth. It can reduce uncertainty where dating is missing, but this study is also a reminder that a rule as old as the 40-week rule deserves to be adjusted to the population it is applied to.
Further reading
Published article: npj Digital Medicine, 5 November 2026 (DOI 10.1038/s41746-026-03453-0), Alliende-Ferrada, Inostroza, Stern, Pérez-Cotapos, Ruz et al. Preprint version: Research Square rs-9999161 (DOI 10.21203/rs.3.rs-9999161/v1). Development study: Patel et al., "Perinatal artificial intelligence in ultrasound (PAIR) study: predicting delivery timing", J Matern Fetal Neonatal Med 2025, DOI 10.1080/14767058.2025.2532099. Reporting checklist: TRIPOD+AI.
Editorial transparency note: this article was produced by Tatakoto's automated editorial pipeline from the preprint version (Research Square, June 2026); the final version published in npj Digital Medicine could not be accessed and some figures may differ slightly. The English, Spanish and Chinese versions are AI-generated. Spotted an error? Contact us.