Building machine learning models to identify placental causes of stillbirth

About This Grant

There are 24,000 stillbirths annually in the United States. In population-level studies, unexplained stillbirths are 30-40%. However, in medical-center based studies with more detailed work-up, the rate is <10%. These studies consistently show that placental histopathology is the most informative datum. This may seem counterintuitive – unlike whole exome sequencing or hemoglobin F flow cytometry, histopathology is widely available. The limitation is the sparsity of human experts. There are around 100 expert perinatal pathologists in the United States, almost all in urban academic medical centers. The situation is unlikely to improve – most projections show the total number of pathologists going down. Our solution is to develop machine learning (ML) models that can make the key placental histopathologic diagnoses in stillbirth. Placental findings may reinforce the significance of information already known, such as evidence of maternal vascular injury in patients with hypertensive disorders of pregnancy. Alternatively, placental findings may be critical, but have no established way of diagnosing them before delivery, such as histiocytic intervillositis. We will test these models in a group of >1500 pregnancies that ended in stillbirth. We will perform detailed abstraction of clinical, obstetric, and laboratory information. Pathologists and maternal-fetal medicine experts will confer on classifying the cause of demise. We will test the importance of placental diagnoses in identifying the anomaly leading to demise and evaluate whether machine learning diagnoses give meaningful information on the causes of demise. To improve generalizability, we will test our model using a set of 500 stillbirths from the Mayo Clinic. Better identifying the cause of demise has immediate and long-term benefits. Understanding what happened is an important part of grieving for patients and families experiencing stillbirth. The benefits are magnified if patients attempt a subsequent pregnancy, since their risk can be better delineated. Knowing the cause of a prior stillbirth can allow personalized treatment. Researchers focusing on specific causes of stillbirth need to identify patients at risk of that cause and to target them for treatment. In the US, stillbirths are reviewed in multidisciplinary conferences. The models developed in this study can improve the work of pathologists in these conferences and, potentially, support decision making in these groups. With sufficient improvement, these models can be piloted in lower-resource settings as full replacement for pathologists or conferences.