Placenta accreta spectrum is one of the most dangerous complications in modern obstetrics. The placenta grows too deeply into the uterine wall and won’t detach after delivery. When surgeons encounter it unexpectedly during a cesarean, massive hemorrhage can follow within minutes. Emergency hysterectomy is often the only option. Maternal death, while less common than it once was, remains a real possibility - a 25-fold increase in mortality risk compared to uncomplicated cesarean delivery.
The problem isn’t that medicine lacks ways to look for it. Ultrasound can catch signs of abnormal placental attachment. The problem is that roughly half of all cases go undiagnosed before delivery. Standard screening depends heavily on the skill of the individual sonographer and the experience of the interpreting physician. Subtle markers get missed. Ambiguous findings get filed as inconclusive.
At the Society for Maternal-Fetal Medicine’s 2026 Pregnancy Meeting in Las Vegas this week, multiple research teams presented AI systems designed to close that gap - and the results, while early, are striking.
The Baylor Study: Zero Missed Cases
The headline result came from Baylor College of Medicine. Dr. Alexandra L. Hammerquist and colleagues trained an AI model on 2D obstetric ultrasound images from 113 patients at risk for PAS who delivered at Texas Children’s Hospital between 2018 and 2025. The patients were scanned at an average gestational age of about 31 weeks.
The model caught every single confirmed case of placenta accreta spectrum. Zero false negatives. It flagged two patients who turned out not to have PAS - false positives - but from a clinical safety perspective, a false alarm that leads to extra monitoring is vastly preferable to a missed diagnosis that leads to uncontrolled bleeding.
“We are hopeful that its use as a screening tool will help decrease PAS-related maternal morbidity and mortality,” Hammerquist said in announcing the findings. The work will be published in the February 2026 issue of Pregnancy.
Multiple Teams, Multiple Approaches
Baylor wasn’t alone. Mount Sinai Health System presented several AI-driven obstetric studies at the same meeting.
Dr. Henri Mitchell Rosenberg showed a preconception machine learning model that mines electronic medical records to identify patients at elevated risk for PAS before they even become pregnant. That model flagged anemia as a novel risk factor - something not currently part of standard PAS risk assessment.
Dr. Tess Cersonsky presented two related studies: one using recurrent neural networks to evaluate how the surgical technique of prior cesarean deliveries influences future PAS risk, and another applying principal component analysis to identify which combinations of clinical factors best predict the condition.
And in a different track at the same meeting, Dr. Jennifer Lam-Rachlin presented an AI tool that flags suspicious findings during fetal cardiac screening for severe congenital heart defects - a different condition, but the same core challenge of making screening more reliable by reducing dependence on individual operator skill.
Separately, a team from King Abdulaziz University published a multimodal deep learning approach on arXiv combining both MRI and ultrasound data. Their model, which fuses a 3D DenseNet121-Vision Transformer for MRI with a 2D ResNet50 for ultrasound, achieved 92.5% accuracy on a dataset of over 1,200 MRI scans and 1,100 ultrasound scans - substantially outperforming either imaging modality alone.
Why Half of Cases Get Missed
Placenta accreta spectrum has become far more common over the past four decades, rising from roughly 1 in 1,250 pregnancies in the 1980s to as many as 1 in 272 today. The primary driver is the rising cesarean delivery rate. Each uterine scar creates a site where the placenta in a subsequent pregnancy can embed abnormally.
Detecting PAS on ultrasound requires looking for specific markers: loss of the clear zone between the placenta and uterine wall, irregular blood vessel patterns, placental tissue bulging through the uterine surface. These signs can be subtle. They require an experienced eye, specific training, and enough time to evaluate them carefully - conditions not always present in busy prenatal clinics.
When PAS is diagnosed in advance, surgical teams can plan for it. They assemble multidisciplinary teams, arrange for blood products, and schedule delivery at centers equipped to handle the complication. Outcomes improve significantly with advance notice. When it’s discovered for the first time on the operating table, clinicians are managing a crisis in real time with limited options.
What This Means
No single study presented at SMFM 2026 is ready for clinical deployment. The Baylor study examined 113 patients at a single institution. The Mount Sinai models haven’t been validated prospectively. The multimodal approach on arXiv used retrospective data. All of these need larger, multi-center trials before they could change clinical practice.
But the convergence is hard to ignore. Multiple independent teams, using different AI architectures and different data sources, are arriving at the same conclusion: automated screening can identify PAS cases that human readers miss. The 50% detection rate that has persisted for years despite improved imaging technology suggests that the bottleneck isn’t the quality of the images - it’s the consistency of their interpretation.
AI screening tools don’t get tired during a long shift. They don’t vary in experience level between hospitals. They evaluate every scan against the same criteria, every time. For a condition where the difference between planned and unplanned management can be the difference between a controlled surgery and an emergency hysterectomy, that consistency could save lives.
The SMFM presentations also suggest the field is moving beyond single-tool approaches. Between preconception risk scoring, automated ultrasound screening, multimodal imaging fusion, and surgical technique analysis, the emerging picture is of an AI-assisted pipeline that could flag high-risk patients before pregnancy, monitor them with more reliable imaging during pregnancy, and help surgical teams prepare when intervention is needed.
The Bigger Picture
This week’s presentations fit a broader pattern in medical AI. The field is shifting from proof-of-concept demonstrations on curated datasets toward clinical validation studies that test models against the messy reality of actual patient care. The University of Michigan’s Prima brain MRI system, published this month in Nature Biomedical Engineering, used the same strategy: train on a health system’s entire digital archive instead of a hand-picked sample, then validate against a full year of real clinical data.
Obstetrics has been slower to adopt AI than some other specialties, partly because the consequences of error are so severe - two patients at once, decisions made under time pressure, conditions that can deteriorate in minutes. The SMFM 2026 presentations suggest that research teams are taking those constraints seriously, designing systems around clinical safety rather than benchmark performance.
The most telling detail in the Baylor study isn’t the zero false negatives. It’s the two false positives. A system tuned for maximum sensitivity, willing to over-flag rather than under-flag, is a system designed by people who understand what happens when PAS surprises a surgical team. That’s clinical thinking, not engineering optimization.
Whether these tools reach clinical practice depends on what happens next: prospective trials, multi-site validation, integration into existing ultrasound workflows, and the regulatory path for AI-assisted diagnostics in obstetrics. But the research direction is clear, and the clinical need is urgent. When a condition this dangerous goes undetected half the time, better screening isn’t optional - it’s overdue.