Machine and Deep Learning Methods for Predicting Preterm Births from EHG Signals

According to the WHO, more than 15 Million babies are born before the full period of the pregnancy. Preterm childbirth is the leading cause of mortality in children less than five years old: in 2015 it was responsible for 1 Million deaths in this age group. While preterm birth is not related to any particular genetic lineage, among the preterm births, only 10 % result in moralities in high-income settings, while the mortality of preterm births in low income countries is as high as 90 %. Detecting preterm cases allows for improved care of the developing foetus via medication, additional care at home for the expectant individual , or even hospitalization for the remaining period of the pregnancy. However, the populations in low income countries also suffer from a lack of access to a systematic healthcare system, leading to preterm births not being detected proactively. Machine learning techniques have a high potential to improve this situation, by providing an automated framework for the detection of critical cases, and thereby providing an alert to the individual, requiring further follow up with medical professionals. In this work, we investigate Machine and Deep Learning ML for EHG Signals techniques for analyzing uterine electrical signals, in particular the Elec-trohysterogram signal, to detect cases of preterm childbirth. A range of machine learning models including Support Vector Machines, Logistic Regression and Decision trees along with Deep Neural Networks such as Convolutional Neural Networks and LSTMs are trained on samples from the TPEHG dataset. We also construct an ensemble of several neural networks and binary classifiers using plurality voting and train them for EHG signal classification. Across all the methods, best results were obtained with the ensemble machine learning classifier, giving an accuracy as high as 98.99 %, sensitivity of 98.3% and specificity of 97.9%.