EHG-Based Preterm Delivery Prediction Algorithm Driven by Transfer Learning

Abstract

Preterm delivery is currently a global concern of maternal and child health, which directly affects infants’ early morbidity, and even death in several severe cases. Therefore, it is particularly important to effectively monitor the uterine contraction of perinatal pregnant women, and to make effective prediction and timely treatment for the possibility of preterm delivery. Electromyography (EHG) signal, an important measurement to predict preterm delivery in clinical practice, shows obvious consistency and correlation with the frequency and intensity of uterine contraction. This paper proposed a deep convolution neural network (DCNN) model based on transfer learning. Specifically, it is based on the VGGNet model, combined with recurrence plot (RP) analysis and transfer learning techniques such as “Fine-tune”, marked as VGGNet19-I3. Optimized with the clinical measured term-preterm EHG database, it showed good auxiliary prediction performances in 78 training and test samples, and achieved a high accuracy of 97.00% in 100 validation samples.