Delivery Date AI

Performance Metrics

Clinical Evidence

The published record is a single paper, and the manufacturer is on it. Patel and colleagues (Journal of Maternal-Fetal and Neonatal Medicine 2025), with Ultrasound AI's founder Robert Bunn as third author, describe the PAIR study at the University of Kentucky: 5,714 women who delivered between 2017 and 2021, 19,940 ultrasound examinations and 877,141 images. Images from 79 per cent of the cohort (4,505 patients) trained a [[deep-learning]] model to estimate days until delivery. The remaining 21 per cent (1,209 patients) was held out, with delivery outcomes blinded from the model by an independent third-party data monitor. Preterm birth rates were 18.4 per cent in training and 18.6 per cent in validation. The first version explained most of the variance in delivery timing for term births (R-squared 0.90) and for term plus spontaneous preterm births together (0.88), but little of it for spontaneous preterm birth alone (0.48). Classified as a preterm-birth test, that version had 39 per cent [[sensitivity]], 93 per cent [[specificity]] and an area under the curve of 0.757. The abstract reports that mean absolute error was similar across the trimesters assessed, without giving the day counts. Three retrainings followed, the last two drawing on a further 1,165,618 images from the same centre to 2023. R-squared for all births rose from 0.85 (version 1) to 0.88 (version 3) and 0.92 (version 4). For spontaneous preterm birth the version 4 mean absolute error was 19.99 days. The manufacturer's clinician FAQ, citing the same study, gives a version 4 mean absolute error of 12.90 days for all births and 10.76 to 15.06 days by trimester. The full paper could not be retrieved for this review, so those figures are not checked. The authors conclude that the model predicts timing until delivery from ultrasound alone and predicts preterm delivery with limited sensitivity. The FDA validation is described only by the manufacturer. Its website states that the pivotal study enrolled 247 women in the cleared population. The model's mean absolute error was 15.22 days (95% CI 13.67 to 16.77), against 36.41 days (32.88 to 39.94) for Hadlock biometric dating, which beat a prespecified 7-day superiority margin (p<0.001). The same page reports the cohort as 62.8 per cent obese, 63.2 per cent White and 31.8 per cent Black, with Asian and Hispanic women under-represented. The design, sites and model version are not stated. The results are not in a peer-reviewed paper, and the De Novo decision summary is not posted. Three gaps shape the grade. The only published study is retrospective, single-centre and company-co-authored. The pivotal figures cannot yet be checked against a public source. And the FDA's special controls require the labelling to state that the output is not intended to predict or assess the risk of preterm birth, which is the outcome most of the paper is about. One De Novo grant and one published study earn a single AurisMark. Two marks would need public prospective evidence in the cleared population, or documented broad uptake, and neither exists yet.

Inside the Auris+ Listing

Regulatory Approvals

Safety Record

The device has been marketable in the US since 11 February 2026. openFDA returned no adverse-event report, recall or enforcement action for the product, the manufacturer or product code SHE as of 13 September 2026 (data release 31 August 2026). The grant letter names the risks the FDA weighed: inappropriate management from a date that is significantly early or late, inconsistent output across acquisition equipment and settings, and users reading the predicted date as if it were the gestational age. The mitigations are clinical performance testing, software validation, labelling and a post-market performance management plan that must monitor real-world performance and tell users what the current performance is.

Intended Use & Indications

Per the FDA De Novo grant letter (DEN250007, 11 February 2026), Delivery Date AI is a post-processing software designed to analyze pregnancy ultrasound images using machine learning techniques. It analyzes raw images to provide healthcare practitioners with a Predicted Delivery Date (PDD) in women without a reliable estimated delivery date. The software estimates the predicted delivery date for a singleton pregnancy and should be used as an aid in clinical judgment, alongside standard methods for assessing gestational age. It processes entire images, which include fetal and maternal characteristics. The intended population is women 18 years of age or older with a singleton pregnancy at or beyond 14 0/7 weeks through 36 6/7 weeks gestation who lack a reliable estimated delivery date because of an unreliable last menstrual period (irregular cycles, spotting mistaken for menses, recent hormonal contraception) or no first-trimester ultrasound, with the pregnancy verified to be within the indicated range by clinical indicators such as fundal height or a fetal heartbeat. It is a prescription device under 21 CFR 801.109.