A Machine Learning Approach for an Early Prediction of Preterm Delivery

Abstract

The preterm birth presents a major cause of the infants' deaths, or the consequent health impairments globally, with an increasing trend of the preterm rate. The enormous global burden on both families and society calls for the preventive and predictive measures. The electrohysterogram (EHG), electrical activity of uterus as measured by surface electrodes, is a noninvasive and affordable tool for effective monitoring of both pregnancy and labour. In this study, the possibility of an early prediction of preterm delivery from the EHG recordings made between 22 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">nd</sup> and 25 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">th</sup> week of the gestation is explored. A set of novel features, including those exploiting signal's nonstationarity, based on the predictive modelling, and empirical mode decomposition, was evaluated on 15min long EHG recordings from the publicly available Term-Preterm EHG (TPEHG) database. On average, Random Forest classifier combined with artificial sampling, tested using 10-fold cross-validation on 322 samples (38 preterm) provided for 99.23% accuracy, with 98.40%sensitivity, and area under curve of 99%. The proposed approach has an additional advantage achieving the classification improvement over shorter, 15min long EHG recordings.