Fetal Electrocardiography and Deep Learning for Prenatal Detection of Congenital Heart Disease

Abstract

Congenital heart disease (CHD) is one of the main problems that can occur during pregnancy. Annually, 300.000 babies die during pregnancy or infancy because of CHD. Early detection of CHD leads to reduced mortality and morbidity, but is hampered by the relatively low detection rates (i.e. <60%) of current CHD screening technology. This detection rate could be improved by complementing echocardiographic screening with assessment of the fetal electrocardiogram (ECG).In this study, the fetal ECG was measured non-invasively, with electrodes on the maternal abdomen, in almost 400 fetuses, 30% of which had known CHD. The fetal ECG measurements were processed to yield a 3-dimensional fetal vectorcardiogram. A deep neural network was trained to classify this fetal vectorcardiogram as either originating from a healthy fetus or CHD. The network was evaluated on a test set of about 100 patients, showing a CHD detection accuracy of 76%. Non-invasive fetal electrocardiography therefore shows clear potential in diagnosis of CHD and should be considered as supplementary technology next to echocardiography.