Its principle is surprisingly precise: analyze the organization of certain cells in the small vessels of the placenta to spot an anomaly that is difficult to identify among thousands of almost identical structures.
Everything seemed to be going well… then the tension can suddenly rise
A pregnancy can take place without warning. The delivery goes normally. Then, a few days later, hypertension appears.
Preeclampsia does not necessarily stop with the birth of the baby. It can also appear after childbirth and is one of the main causes of pregnancy-related death.
The placenta could then provide valuable clues. Its analysis sometimes makes it possible to find traces of complications that occurred during pregnancy. But in practice, not all placentas are examined in detail.
In the United States, less than 20% would be analyzed after deliveryparticularly due to the lack of pathologists specialized in perinatal pathology.
This is precisely the problem that researchers at Carnegie Mellon University and UPMC are seeking to solve. Their artificial intelligence, described in npj Digital Medicineautomatically analyzes digital images of the placenta and flags cases requiring further examination.
To understand what she is looking for, we have to go down to the scale of a few cells.
Among 4,000 ships, the AI searches for those that tell a different story
The challenge is considerable. Researchers must be able to spot only a few abnormal vessels among 3,000 to 4,000 ships who look alike.
Mangalam Sahai, first author of the study, compares this task to the famous observation game:
“Examining placentas is a bit like the game Where’s Waldo?. We’re looking for five diseased blood vessels in a sea of 3,000 to 4,000 vessels that look very similar“.
The anomaly sought is called decidual vasculopathy. This damage to the maternal blood vessels of the placenta is frequently associated with postpartum preeclampsia.
The researchers were interested in two cell populations: extravillous trophoblast cells (EVT)which normally participate in the remodeling of maternal arteries, and red blood cells.
In a healthy vessel, these cells follow a particular organization. In a vessel affected by decidual vasculopathy, the distribution changes: EVT cells are more present in the wall.
The algorithm analyzes this spatial organization based on the optical properties of the colored cells. It then transforms this data into standardized values and calculates a “morphological separation score”.
So the AI doesn’t just say that a ship looks sick. She seeks to measure what distinguishes its cellular organization from a normal vessel.
Help for doctors, not a replacement
In practice, the algorithm scans images and flags unusual samples. The specialist pathologist can then focus on suspected cases.
The issue is particularly important in establishments which do not have an expert in perinatal pathology.
“Human intervention cannot be replaced“, insists Jon Cagan, professor of mechanical engineering. “AI models will enable pathologists to diagnose more cases of preeclampsia throughout the day, helping to make clinical decisions more efficient and accessible.”.
This place left to the doctor is essential. The algorithm can spot a signal, but it does not replace the interpretation of a patient’s entire medical record.
The researchers also highlight another interest:
AI can explain what it observes.
“We are developing a biologically relevant AI model that can contribute to diagnostics.”said Phil LeDuc, professor of mechanical engineering. “This is essential to ensure reliable diagnostics. I am convinced that in the future, models like ours will improve the lives not only of patients with preeclampsia, but also of those suffering from cancer and brain diseases, for which different biomarkers are early indicators.”.
The placenta, a memory of pregnancy to be better exploited
This approach is part of a broader movement. Other teams are already working on artificial intelligence models capable of estimating the risk of preeclampsia from the medical data of pregnant women. A team from Weill Cornell Medicine notably studied nearly
59,000 pregnancy files.
The tool developed by Carnegie Mellon and UPMC intervenes differently: after childbirthusing the placenta as a sort of biological memory of the pregnancy. This avenue could be particularly interesting when preeclampsia appears after returning home. A woman can leave the maternity ward without obvious symptoms and then develop hypertension requiring rapid treatment.
But caution remains essential. This AI is not currently a tool that can alone diagnose preeclampsia or predict with certainty that a woman will develop a complication. Its interest is rather to give doctors
an extra lookcapable of rapidly scanning thousands of cells and vessels in search of a precise signal.
The placenta, which is often considered useless once the birth is over, could thus become a real part of the medical file. And artificial intelligence could help find clues that the human eye, faced with the immensity of the tissues to examine, risks missing. For patients, the issue is ultimately very concrete: better understand what happened during pregnancy and not let an important signal disappear when they leave the maternity ward.