Abstract

Traditionally, pregnant women in the United States who participate in prenatal care have been scheduled for approximately 14-16 prenatal visits, which is the schedule recommended by the American College of Obstetricians and Gynecologists (OBS/GYN). The closer the delivery day, the more frequent the inspections. Clinically, doctors mostly used ultrasound examination to estimate the time of fertilization, weeks gestation, and the expected date of childbirth, which has always played a key role about how often women need to have a checkup when delivery time approaches. Therefore, a accurate enough expected date (week-level accuracy) not only affects prenatal visits, but also reduce the psychological stress and anxiety for the women, and help women and their family make schedule about applications for maternity leave, tocolysis leave, or any preparation for baby. Our contribution is a high prediction accuracy according to DNN methodology. We compared dense labeling after DNN model, and sequence-based labeling after our application labeling. We got a excellent accuracy over 95% by seeing F1-score of six classes. A consistency we observed is that different classes shows different level of energy in frequency domain, which means, in clinical, doctors always only focus on uterine contraction, but in fact, the features in frequency domain are truly influential the time of delivery. Finally we conclude that we can still obtain a high enough accuracy (over 90%) if we only use two channels instead of all 16 channels. And another observation was that the accuracy is not concerned with the positions of electrodes. So all these conclusions may be helpful in future works in field of OBS/GYN.

Introduction Video