AI helps pathologists find signs of preeclampsia, advancing diagnosis and treatment

Preeclampsia is one of the leading causes of pregnancy-related death. Even after a healthy delivery, mom and baby can go home only to show signs of postpartum hypertensive disorder days or weeks later.

Examining the placenta after delivery can provide important clues about complications that occurred during pregnancy and identify future risk. One such clue is decidual vasculopathy, a disease of the material blood vessels in the placenta that is commonly associated with postpartum preeclampsia. However, there is a profound shortage of perinatal pathologists trained to conduct routine placenta screenings, meaning that today less than 20% of placentas are screened in the United States after delivery. 

To streamline this process with the hope that one day all placentas are reviewed by pathologists for signs of disease, researchers in the Department of Mechanical Engineering at Carnegie Mellon University and the Department of Pathology at UPMC have developed a machine learning algorithm that can quickly identify placentas with notable biomarkers and flag cases that may require further review from a specialized pathologist.

“Screening placentas is a bit like playing Where’s Waldo,” said Mangalam Sahai, first author of the research paper published in npj Digital Medicine, “We are looking for five diseased blood vessels in a sea of three to four thousand very similar looking vessels.”

One of the biological differences pathologists look for involves extravillous trophoblast cells (EVT), cells that help remodel the maternal blood vessels to make sure that enough blood is reaching the placenta and the developing fetus. In vessels affected by decidual vasculopathy, the organization of these cells relative to red blood cells changes. The researchers’ first step was, therefore, to use stain-dependent light transmittance properties of EVT and red blood cells within images of placental vessels to identify their spatial organization.

This is a bit like comparing two types of trail mix: both contain the same ingredients, but in different proportions. By looking at the relative amount of each ingredient, you can distinguish one mix from the other. Similarly, the model looks at the relative distribution of EVTs and red blood cells within a vessel to distinguish healthy from diseased vessels. The algorithm converts those patterns into a standardized score that indicates how typical or unusual the cell organization is. Rather than simply labeling a vessel as “diseased” or “healthy,” the team’s model stands out because of a third, unique step that calculates what the researchers call a morphology separation score. This score compares the relative distributions of EVTs and red blood cells within each vessel. Healthy vessels tend to have fewer EVTs relative to red blood cells while diseased vessels have more EVTs. These differences cause the vessels to separate into two distinct groups.

The model’s novel ability to not only determine whether or not there are signs of disease, but to also explain why it came to that conclusion, is critical to deploying AI in health care. Explainability is very important in AI, rather than relying on an unexplained “black box diagnosis,” and thus the model can point pathologists to a measurable biological difference.

I believe that in the future, models like ours will improve the lives of not only patients with preeclampsia, but those with cancer and brain related diseases where different biomarkers are an early indicator.

Phil LeDuc, Professor, Mechanical Engineering

“We are introducing a biologically relevant AI model that can help explain diagnosis,” said Phil LeDuc, professor of mechanical engineering. “This is very important in building trustworthiness in diagnosis. I believe that in the future, models like ours will improve the lives of not only patients with preeclampsia but those with cancer and brain-related diseases where different biomarkers are an early indicator.”

“Human intervention cannot be replaced,” emphasized Jon Cagan, professor of mechanical engineering. “AI models will enable pathologists to diagnose more preeclampsia patients throughout the day to help make clinical decisions a more efficient, more democratized process.”