AI Model Detects Placenta Accreta With No False Negatives in 113-Patient Test

Placenta accreta spectrum is one of obstetrics' more treacherous diagnoses. The placenta embeds abnormally deep into the uterine wall - sometimes through it entirely - and the result at delivery can be catastrophic hemorrhage, organ failure, and death. The condition's incidence is rising in the United States, driven by increasing rates of cesarean delivery, which creates the uterine scarring that predisposes the abnormal attachment.

The problem is not just severity. It's detection. Current screening, which combines risk-factor assessment with ultrasound imaging, fails to identify approximately half of all cases before delivery. Women arrive in operating rooms where surgeons encounter a complication they weren't prepared for.

What the AI model was asked to do

A team from Baylor College of Medicine built a deep-learning model trained to analyze standard 2D obstetric ultrasound images - the same imaging used in routine prenatal care. They then applied it retrospectively to ultrasound data from 113 patients who had been identified as high-risk for placenta accreta spectrum and who delivered at Texas Children's Hospital between 2018 and 2025.

The mean gestational age at the time of the ultrasound scans was 30.89 weeks, plus or minus 3.67 weeks - a window in the third trimester when there is still time to plan a delivery strategy if accreta is confirmed.

The model's performance on this cohort was striking. It correctly identified every confirmed case of placenta accreta spectrum. False negatives - cases the AI missed - numbered zero. There were two false positives, meaning two patients flagged as having accreta who did not.

Why the false negative rate matters most

In clinical terms, a false negative in this context is the dangerous error. A missed diagnosis means an unprepared surgical team, potentially no blood products staged, and no specialist maternal-fetal medicine surgeon scrubbed in. Two false positives, by contrast, trigger extra precautions that turn out to be unnecessary - a manageable cost when weighed against the alternative.

"We are hopeful that its use as a screening tool will help decrease PAS-related maternal morbidity and mortality," said Alexandra L. Hammerquist, a maternal-fetal medicine fellow at Baylor and one of the study's lead researchers.

What the data cannot yet show

The study has important limitations that require clear acknowledgment. The 113-patient cohort was retrospective and drawn from a single high-risk center - not a random sample of the general obstetric population. All patients were already flagged as at-risk before the AI assessed their scans, which means the model has not been tested on low-risk pregnancies where accreta is less prevalent but still possible.

Retrospective studies of AI diagnostic tools can reflect overfitting - where a model learns patterns specific to one institution's imaging equipment, technician technique, or patient population that don't generalize elsewhere. Prospective validation across multiple centers, with varied equipment and broader patient demographics, is the necessary next step before clinical deployment.

The case count of 113 is relatively modest for training and testing a medical AI. The performance figures - zero false negatives, two false positives - are encouraging, but confidence intervals around those numbers would narrow considerably with a larger validation cohort.

A condition worth the investment in better screening

Placenta accreta spectrum affected an estimated 1 in 272 deliveries in recent U.S. data, and that rate has been climbing alongside cesarean delivery rates for decades. Each additional cesarean a patient undergoes roughly doubles the probability of accreta in a subsequent pregnancy. As the obstetric population carries more uterine scars, the need for reliable prenatal detection grows correspondingly.

The research was presented at the Society for Maternal-Fetal Medicine 2026 Pregnancy Meeting in Las Vegas, and the abstract is slated for publication in the February 2026 issue of PREGNANCY, the Society's peer-reviewed journal.