Abstract:
Current Critical Congenital Heart Disease (CCHD) screening uses pulse oximetrymeasured
oxygen saturation (SpO2) and fails to detect an estimated 900 newborns
annually in the US. Recent studies have shown additional pulse oximetry
measurements such as heart rate and perfusion index allow for improved detection of
CCHD through use of Machine Learning (ML), especially for those cases with systemic
blood flow obstruction such as Coarctation of the Aorta (CoA). However, acquiring pulse
oximetry measurements on newborns is a challenge, exacerbated by
Photoplethysmography (PPG) signals being prone to motion artifacts. We developed a
full pipeline for measuring, filtering (i.e., automatically removing artifacts), and analyzing
newborn's pulse oximetry data based on the Pi-Top™ and Nonin©WristOx2™3150
devices, and subsequently training a random forest classifier for early detection of
CCHD. Our pipeline includes a ML based CCHD detection model using radiofemoral
delay and PPG characteristics in addition to SpO2, heart rate, and perfusion index. Our
model achieves 100% Specificity and 93.33% Sensitivity.
- Tags
-
