Can ultrasound predict
when delivery
will happen?

The PAIR study: evidence, limitations, and clinical meaning

Chukwuma Onyeije, MD · Atlanta Perinatal Associates

Two questions. Different endpoints.

PAIR does not validate assigning or revising an EDD in a pregnancy with uncertain dates.

A large image dataset.
One institutional setting.

Retrospective cohort · University of Kentucky · 2017–2021

Random allocation by patient; delivery outcomes in the held-out set were concealed from the AI and developer through an independent monitor.

Images in. Days until delivery out.

Missing gestational age at delivery led to exclusion. Stillbirth outcomes were unavailable.

Track the version with the result.

The methods state that continuous days-until-delivery prediction was explored after the original preterm-versus-term classifier proved feasible.

Classification results are reported for V1/V2; the principal updated delivery-timing results are V3/V4.

Delivery timing improved with retraining

R² describes variation explained across observations. It does not mean that 92% of patients delivered on the predicted date.

Prediction error remains substantial.

V4 mean absolute error (MAE), in days · lower is better

About 20 days of average absolute error for spontaneous preterm birth. MAE is not a patient-specific prediction interval.

Predictions are per examination, not one independent observation per patient.

Performance across trimesters

1,714 examinations; trimester sample sizes: 327 / 791 / 596. Overlapping confidence intervals do not establish equivalence.

These stratified analyses evaluated time until delivery. They do not establish accuracy for gestational age or EDD assignment in patients first scanned late in pregnancy.

Preterm birth classification

At 40% sensitivity, roughly 6 in 10 preterm births are missed at the reported operating point.

Preterm birth prevalence in the validation set was 18.6%. Sensitivity remained low, so a negative prediction cannot reliably exclude preterm birth.

What strengthens the study?

• Held-out patients had delivery outcomes blinded from the AI by an independent monitor.
• The authors reported both regression and classification metrics, with separate preterm subgroups.
• Iterative retraining allowed examination of changes in performance as the image dataset expanded.

What remains unproven?

Implications for pregnancies with uncertain dates

A late presenting patient needs an estimated gestational age and EDD for clinical decisions. PAIR predicts a different endpoint: days until actual delivery.

Before considering a dating use case, ask the vendor for the exact intended use, evidence for gestational age estimation in second and third trimester presentations, external validation, error distributions, and handling of uncertainty.

Keep the established EDD and any predicted delivery date in distinct fields during evaluation.

A focused APA evaluation

1. Define the eligible late presenting population and document how each current EDD was assigned.
2. Run the model in observation mode and record its output separately from the clinical EDD.
3. Compare predictions with observed delivery timing and with simple clinical baselines.
4. Audit performance by scan trimester, spontaneous versus indicated delivery, and key patient subgroups.
5. Require prospective evidence for a separate gestational dating claim before considering a change to the EDD workflow.

Promising signal.
Unresolved clinical utility.

The next step is independent, prospective validation against clinical baselines, with uncertainty and patient benefit measured explicitly.

Gestational dating requires separate evidence.

Source and disclosures

Patel N, O’Brien J, Bunn R, Schanbacher B, Bauer J, Lam GK. Perinatal artificial intelligence in ultrasound (PAIR) study: predicting delivery timing. J Matern Fetal Neonatal Med. 2025;38(1):2532099.

doi:10.1080/14767058.2025.2532099 ↗