Improving the prediction of pregnancy complications using artificial intelligence
One in ten pregnant women experience signs of preeclampsia – elevated blood pressure and increased protein in the urine – at some point in their pregnancy. Only a few of these women actually go on to have serious complications. But since these can lead to the death of the mother and baby, the pregnant women are often admitted to the hospital just to be safe, though in most cases this turns out in retrospect to have been unnecessary. Scientists at Charité – Universitätsmedizin Berlin have collected data from 1,647 preeclampsia patients and used machine learning methods to create a model that can better predict the probability of serious complications.
Preeclampsia, formerly known as toxemia of pregnancy, has a variety of signs and symptoms: The blood pressure rises suddenly and quickly; excessive protein is found in the urine; fluid accumulates in the arms or legs, a condition known as edema; those affected experience dizziness, headaches, nausea or upper abdominal pain. “About 10 to 15 percent of all pregnant women have a clinical suspicion of preeclampsia at least once during their pregnancy,” reports Professor Stefan Verlohren, senior physician in the Department of Obstetrics at Campus Charité Mitte. “Yet only two to five percent actually develop the disease, and serious complications occur in only a small proportion of these women.”
“More accurate predictions simply cannot be achieved through blood pressure and urine protein measurement alone. And although we’ve recently introduced new biomarkers like sFlt-1 and PIGF, we wanted to explore the hypothesis of using machine learning algorithms to jointly analyze multiple test results in order to better predict who is at risk and who is not.”
1,647 patient data sets were evaluated
First, the scientists led by Verlohren compiled a database of real treatment data from a total of 1,647 high-risk preeclampsia patients who presented at Charité’s Department of Obstetrics between 2010 and 2019. With the help of these data sets, the team used machine learning methods or artificial intelligence (AI) to calculate how likely it is that complications would occur. They included a total of 114 different features in the evaluation.
“After training the algorithms with the clinical data sets, we compared the prediction accuracy of our AI model with the traditional prediction method based on clinical parameters,” Leon Schmidt, who is conducting his doctoral thesis in Verlohrens group, reports. “Our algorithms were clearly superior to the traditional predictions based on blood pressure, proteinuria and the sFlt-1/PlGF ratio. Of particular note was that the positive predictive value – i.e., that a complication will actually occur – was double as accurate.”
AI cannot replace doctors – but it can lend a helping hand
The Digital Health Accelerator Program of the Berlin Institute of Health at Charité (BIH) is supporting further product development so that a solution based on the newly created algorithm can be rapidly translated into broad clinical practice, thus making it widely available to patients and doctors in the future. However, prospective studies must first confirm the validity of the algorithm. “AI of course cannot replace physicians,” Verlohren, a physician himself, is convinced. “But when it comes to considering how preeclampsia treatment should proceed, AI certainly provides a better basis for making decisions. This could prevent potentially life-threatening complications for both mother and child.”
Contact for scientific information:
Dr. Stefan Verlohren
Originalveröffentlichung: Am J Obstet Gynecol. 2022 Jan 31:S0002-9378(22)00050-3.