pregnant woman holding ultrasound results

Photo by Christin Noelle from Unsplash

Fair Tests Cut AI’s Due Date Edge to Under a Day

In A Nutshell

  • An AI ultrasound tool predicted delivery dates nearly five days closer than standard due dates in a Chilean study of 3,788 exams.
  • Most of that edge came from Chilean pregnancies running shorter than 280 days, leaving under a day once that was counted.
  • A separate preterm birth risk score outperformed gestational age alone in past records, though it over-forecast risk by about a fifth.
  • No patient was managed with the tool, so real-world testing must come first.

An artificial intelligence tool that reads ultrasound images landed nearly five days closer to the actual delivery date than the standard due date did, on average. Most of that edge traced back to a simple fact: pregnancies in the Chilean hospitals studied tend to end earlier than the textbook 280 days. Once comparisons accounted for that, the AI’s lead shrank to less than a day, too small to change any medical decision.

Standard due dates assume a 280-day pregnancy, counted from the last menstrual period or fixed by an early ultrasound measurement. For women who reach care late, with no reliable period record or early scan, that estimate gets shakier. The AI tool promised a workaround: reading ultrasound images directly, with no menstrual history or measurements needed.

A team of Chilean researchers put a commercial version of that tool to the test on thousands of real exams. On raw numbers, it looked like a major upgrade, missing the delivery date by 6.9 days on average versus 11.6 days for the standard due date. Where the tool showed real strength was in something arguably more useful than a slightly better due date: identifying which pregnancies were at risk of giving birth prematurely.

Chilean Hospitals Tested the AI Ultrasound Tool on 3,788 Exams

Writing in npj Digital Medicine, researchers described ultrasound exams performed between 2021 and 2025 at four public hospitals in Chile. After twin pregnancies and records with missing or unmatched information were set aside, the dataset held 3,788 exams from 2,550 pregnancies. About a third of the women were migrants, most commonly from Venezuela and Haiti.

That mix mattered. Haitian women were twice as likely as Chilean women to arrive for care without an early scan. Preterm birth, meaning delivery before 37 weeks, was far more common among pregnancies without an early scan, at 15% versus 6.3% for those with a reliable one. A wrong due date can affect when to induce labor and when to give medications that help a baby’s lungs mature.

Nobody retrained the tool for Chile. It analyzed the stored images as it would anywhere else, and no patient was managed using its output.

AI Ultrasound Due Dates_ Small Reality
AI ultrasound may be better at flagging premature birth risk than predicting due dates, a study of 3,788 exams found. (Image by StudyFinds)

AI Due Date Edge Shrank to Under a Day Once Shorter Pregnancies Were Counted

Births in this group arrived about 11 days before the 40-week mark on average, so the standard due date pointed too late. A tool trained on actual birth dates, rather than the 280-day rule, avoids that built-in error, which explains much of its head start.

Researchers rebuilt the comparison using estimates that already reflected the local pattern. Against those, the AI’s edge dropped to between half a day and three-quarters of a day. The authors said an advantage that small “is not a clinical quantity, whatever its formal significance,” meaning it is too small to guide care even if a statistical test calls it real.

A closer look showed where the AI did pull ahead. Among full-term pregnancies, a simple formula based only on how far along the woman was performed about as well as the AI. Among pregnancies that ended early, that formula missed by an average of 18 to 28 days, against 6 to 12 for the AI.

AI Ultrasound Risk Score Spotted Preterm Births Far Better Than Chance

Alongside its date prediction, the system produces a separate risk score estimating the chance of preterm birth. The main test used one exam per pregnancy, taken between 20 and 28 weeks, across 928 pregnancies, 77 of which ended in preterm birth. On a scale where 0.5 is a coin flip and 1.0 is perfect, the AI’s score reached 0.86. Gestational age alone, meaning how far along the pregnancy was, scored 0.52, close to chance.

Outside that strict test, the score also beat a baseline built from medical-record diagnoses and routine measurements from the same exam, though by a smaller margin. It worked at least as well when labor began on its own as when doctors triggered delivery, which argues against the score merely picking up on a clinician’s worry.

Still, the score over-forecast preterm birth by about a fifth across the whole group, so its probabilities cannot be trusted at face value without adjustment. That overestimate showed up at a similar level across nationality groups, with no clear evidence that migrant women were disproportionately affected. Smaller groups limit that conclusion, though: just 19 exams from Haitian women were followed by a preterm birth.

Headline accuracy numbers can flatter a tool, and this one is a case in point. The more promising result, flagging the risk of premature birth, could someday matter most for women who reach care without an early scan, but that remains a hypothesis. Authors said real-world testing must come first, and that the tool should not guide care based on this paper alone.


Disclaimer: This article is for informational purposes only and is not medical advice. The research described is preliminary, and nothing here should guide pregnancy care decisions. Anyone with questions about a due date or pregnancy health should speak with a qualified health care provider.


Paper Notes

Limitations

Researchers analyzed past exams and outcomes rather than using the tool during real care, so no one knows whether acting on its scores would change results for patients. The tool is a proprietary commercial system, so the team could not inspect which visual patterns it relies on, and could only partly rule out that it reads quirks in the images rather than features of the pregnancies. The authors labeled the preterm risk analyses exploratory, meaning they flag patterns to test rather than prove anything, and some groups were small, including Haitian and other migrant women with preterm births, too few to confirm equal performance across groups. All data came from one national health system with one care protocol and pregnancies shorter than 280 days, so results may not carry over to other countries. Cervical length, an established preterm predictor, was recorded in only 3.5% of exams, and 27 pregnancies ending in stillbirth or newborn death were too few to say how the tool behaves in those cases. Findings apply to one version of the software, which the developer may update.

Funding and Disclosures

Authors reported no funding for this work. First author Rodrigo Alliende-Ferrada disclosed a financial relationship as a strategic advisor to Ultrasound AI, Inc., the company that makes the tool, in a role concerning future gynecologic-oncology imaging projects. The authors state that no benefit under that role arises from or depends on this study. The company’s only contribution was unpaid access to its web platform for uploading de-identified image files. It had no role in the design, conduct, analysis, interpretation, writing or submission of the study, and it received no clinical data or delivery outcomes. Co-author Andrés Ruz, who has no relationship with the company, independently verified the analyses.

Publication Details

Rodrigo Alliende-Ferrada (Hospital Padre Hurtado and Hospital Santiago Oriente Dr. Luis Tisné Brousse, Santiago, Chile; corresponding author), Ignacia Inostroza, Jocelyn Stern, Cristobal Pérez-Cotapos and Andrés Ruz wrote “External validation of an image-only ultrasound model for delivery-date prediction in four Chilean public hospitals.” It was published in npj Digital Medicine, in partnership with Seoul National University Bundang Hospital, as an Article in Press. The paper was received June 11, 2026, accepted October 18, 2026, and published online November 5, 2026. DOI: https://doi.org/10.1038/s41746-026-03453-0

About StudyFinds Analysis

Called "brilliant," "fantastic," and "spot on" by scientists and researchers, our acclaimed StudyFinds Analysis articles are created using an exclusive AI-based model with complete human oversight by the StudyFinds Editorial Team. For these articles, we use an unparalleled LLM process across multiple systems to analyze entire journal papers, extract data, and create accurate, accessible content. Our writing and editing team proofreads and polishes each and every article before publishing. With recent studies showing that artificial intelligence can interpret scientific research as well as (or even better) than field experts and specialists, StudyFinds was among the earliest to adopt and test this technology before approving its widespread use on our site. We stand by our practice and continuously update our processes to ensure the very highest level of accuracy. Read our AI Policy (link below) for more information.

Our Editorial Process

StudyFinds publishes digestible, agenda-free, transparent research summaries that are intended to inform the reader as well as stir civil, educated debate. We do not agree nor disagree with any of the studies we post, rather, we encourage our readers to debate the veracity of the findings themselves. All articles published on StudyFinds are vetted by our editors prior to publication and include links back to the source or corresponding journal article, if possible.

Our Editorial Team

Steve Fink

Editor-in-Chief

John Anderer

Associate Editor