FDA Clears AI Tool That Predicts Delivery Date Directly From Ultrasound Images
Predicting when a baby will be born is harder than it sounds. The standard approach - using the date of the last menstrual period, adjusted by early ultrasound measurements of fetal size - works well when periods are regular and first-trimester scans are available. For a significant number of pregnancies, those conditions do not hold. Irregular cycles, late entry to prenatal care, uncertain last menstrual period dates, and in vitro fertilization can all make traditional dating unreliable. In those cases, obstetricians face consequential decisions - when to schedule delivery, when to intervene for post-term pregnancy, whether a fetus is appropriately grown for its gestational age - without a reliable anchor.
A deep-learning system developed by Ultrasound AI has now received FDA De Novo clearance to address that gap. Called Delivery Date AI, the tool analyzes standard ultrasound images and produces a Predicted Delivery Date without relying on biometric measurements taken from the images or on last menstrual period records. The FDA decision marks the first regulatory authorization for an AI system that predicts delivery timing from image data alone rather than from manually extracted fetal measurements.
How the System Works
Delivery Date AI is built on an ensemble of deep-learning neural networks trained on millions of de-identified ultrasound images collected across diverse patient populations and clinical settings. Rather than instructing the model to measure specific fetal structures - crown-rump length, biparietal diameter, femur length - the training approach allowed the networks to identify any features in the full image that correlate with delivery timing, including fetal characteristics, maternal characteristics, and patterns that may not correspond to any single anatomical measurement.
The system is cloud-based. After a clinician uploads an ultrasound image from a standard machine, the analysis runs remotely and returns a Predicted Delivery Date in seconds. Compatible with most existing ultrasound equipment, it requires no hardware modification and minimal staff training.
The Evidence Base
The FDA clearance was supported by data from the PAIR (Perinatal Artificial Intelligence in Ultrasound Research) Study, conducted in collaboration with the University of Kentucky and published in The Journal of Maternal-Fetal and Neonatal Medicine. The study enrolled more than 5,700 patients and evaluated the system's accuracy in predicting days to delivery.
The reported accuracy was an R-squared value of 0.92 - meaning that 92 percent of the variance in actual delivery timing was explained by the model's predictions. For context, a perfect predictor would achieve an R-squared of 1.0, while a model with no predictive value would score near zero. The 0.92 figure represents a high level of accuracy for a biological prediction task, though it also means approximately 8 percent of the variance in delivery timing was not captured by the model and likely reflects individual biological variation that imaging cannot fully characterize.
Target Use Case and Limitations
The FDA cleared Delivery Date AI as an adjunctive tool - one designed to support, not replace, clinical judgment. It is specifically indicated for pregnancies where traditional dating methods are unreliable. The company is not positioning the system as a replacement for standard gestational-age dating in uncomplicated pregnancies where last menstrual period records and early ultrasound measurements are available and concordant.
The system's performance across all subgroups represented in the training data, including patients with high-risk pregnancies, multiple gestations, or fetal growth abnormalities, has not yet been reported in full detail in the published literature. Real-world performance may differ from the controlled conditions of a study cohort, and ongoing post-market evaluation will be essential to characterize the tool's behavior across the full breadth of obstetric practice.