Bio-process inspired characterization of pregnancy evolution using entropy and its application in preterm birth detection

Abstract

The lack of effective method for an early diagnosis of preterm births makes it a public health problem world-widely. The relationship between the uterine contraction and the underlying electrical muscle cells makes the machine-learning-based Eelctrohysterogram (EHG) signal processing an ideal direction for the development of methods for preterm births. Inspired by the observation of dynamical changes of the uterus experiences throughout the whole pregnancy, we used entropy features extracted from the time–frequency expansion of the original EHG signal to characterize the uterine activities. These entropy features were then rescaled by the gestational age at recording (recording time) to characterize the evolution speed of the uterus toward delivery. By selecting out the most relevant frequency components using the principle components analysis (PCA), the Gaussian Naive Bayes (GNB) classifier trained and evaluated with samples prepared under the Partition-Synthesis oversampling scheme gives average preterm births prediction accuracy, sensitivity, specificity and AUC values as high as 0.75, 0.84, 0.66, and 0.84, respectively.