What happened
Researchers at Carnegie Mellon University and UPMC developed an unsupervised machine learning algorithm to identify decidual vasculopathy in placental tissue, a key indicator of postpartum preeclampsia. The tool analyzes the spatial organization of extravillous trophoblast cells and red blood cells to flag cases requiring specialist review, aiming to increase screening rates in the US where fewer than 20% of placentas are currently examined.
Researchers in the Department of Mechanical Engineering at Carnegie Mellon University and the Department of Pathology at UPMC developed a machine learning algorithm to streamline placental screening for preeclampsia. The study, published in npj Digital Medicine, addresses the challenge that less than 20% of placentas in the United States are screened after delivery due to a profound shortage of trained perinatal pathologists.
The algorithm focuses on identifying decidual vasculopathy, a disease of maternal blood vessels in the placenta associated with postpartum preeclampsia. The model uses stain-dependent light transmittance properties to analyze the spatial organization of extravillous trophoblast cells (EVT) and red blood cells within placental vessels. Healthy vessels have fewer EVTs relative to red blood cells, while diseased vessels have more, creating distinct patterns that the AI can detect.
A key feature of the model is its 'morphology separation score,' which quantifies how typical or unusual the cell organization is. This approach allows the AI to not only flag potential disease but also explain the biological basis for its conclusion, a critical factor for building trust in clinical AI applications. The researchers describe the process as similar to finding specific diseased vessels among thousands of similar-looking healthy ones.
Source details: medicalxpress.com ↗
Why it matters
Preeclampsia is a leading cause of pregnancy-related death, yet diagnostic capacity is severely limited by a shortage of perinatal pathologists. This AI tool offers a scalable solution to automate initial screening, potentially allowing pathologists to review more cases and detect postpartum complications earlier. The model’s focus on explainability, using a 'morphology separation score' to highlight biological differences, addresses a major barrier to clinical adoption by providing transparent reasoning rather than a black-box output.
Preeclampsia remains a leading cause of pregnancy-related death, and postpartum complications can emerge days or weeks after delivery. The current diagnostic bottleneck is human capacity; without automated screening, many at-risk patients may not receive timely identification of placental abnormalities.
The development of an explainable AI model is significant for healthcare adoption. By providing a measurable biological difference rather than a black-box diagnosis, the tool aims to facilitate trust among pathologists. This transparency is essential for integrating AI into high-stakes medical decisions where accountability is paramount.
The potential impact extends beyond preeclampsia. The researchers suggest that similar biomarker-based AI models could be applied to other conditions, such as cancer and brain-related diseases, where early detection of specific cellular changes is critical for treatment outcomes.
What to watch next
Clinical validation studies to determine the algorithm's sensitivity and specificity in real-world hospital settings. Regulatory pathways for FDA clearance of the software as a medical device. Integration of the tool into existing hospital pathology workflows and its impact on the overall rate of placental screening.
The next critical step is independent clinical validation to confirm the algorithm's accuracy in diverse patient populations. While the study demonstrates technical feasibility, real-world performance in hospital settings will determine its clinical utility.
Regulatory approval will be necessary before the tool can be widely deployed. The FDA's framework for AI-based medical devices will likely require rigorous testing to ensure safety and efficacy, which may take several years.
Healthcare systems will need to evaluate the cost-effectiveness of implementing such screening tools. If the AI can significantly increase the volume of placentas reviewed by pathologists, it could lead to earlier interventions and improved maternal health outcomes.