PREDICTION OF LABOR TYPES THROUGH ELECTROHYSTEROGRAPHY SIGNAL ANALYSIS

Abstract

In 2019, around 900,000 children died due to complications related to preterm delivery (PTD), defined as births before 37 weeks of gestation. PTD is one of the top five causes of infant mortality, alongside maternal pregnancy complications, and can require induced or cesarean deliveries. Over the past decade, the rates of induced and cesarean deliveries have risen without clear justification, and outcomes can worsen when an induced delivery results in a cesarean. The decision between labor induction and cesarean delivery involves numerous factors and presents a significant challenge in obstetrics.

Accurately distinguishing between different types of labor is crucial due to the significant contribution of maternal complications to infant mortality. Various tools exist for monitoring and detecting early signs of labor, such as cervical length assessment, intrauterine pressure monitoring, biochemical markers, and contraction frequency. However, no single technique reliably predicts preterm delivery.

A promising new approach, Electrohysterography (EHG), has emerged for accurately predicting both preterm and term births. Studies utilizing machine learning and signal feature analysis suggest that EHG can effectively distinguish between preterm and term births. Many of these studies use the Term Preterm EHG Database (TPEHGDB) and employ various machine learning algorithms and features. Recently, another database, the Induced Cesarean EHG DataSet (ICEHGDS), from the same data pool as TPEHGDB, has been published. However, there is a lack of studies distinguishing between induced and cesarean deliveries from this database.

This project aims to address current maternal and neonatal mortality rates by accurately distinguishing between natural preterm and term births using TPEHGDB, between induced and cesarean term births using ICEHGDS, and by analyzing all four categories with both databases, recorded with the same experimental setup.

Three approaches are implemented in the machine learning algorithms: starting with signal features alone, then analyzing performance by adding sociodemographic features, and finally refining the algorithms by focusing on the most relevant features. For the four types of birth analysis, the project initially employs signal and sociodemographic features, then incorporates the most significant features. This methodology aims to determine the best classification performance. The algorithms used include K-Nearest Neighbors (KNN), Support Vector Machines (SVM) with Gaussian, linear, and polynomial kernels, Decision Trees (DT), and ensemble methods like Random Forest (RF) and AdaBoost. A 5-fold cross-validation evaluates and compares the results. The metrics studied are accuracy, sensitivity, specificity, and F1 score. For the four types of birth analysis, confusion matrices are included.

In distinguishing between premature and term or induced and cesarean, SVM with a Gaussian kernel performs best using the most relevant sociodemographic and signal features, achieving 98.4% accuracy (SD=1.4%), 98% sensitivity (SD=2.4%), 99.0% specificity (SD=2.2%), and an F1 score of 98.4% (SD=1.5%). In distinguishing between cesarean and induced, SVM with a Gaussian kernel also performs best with 95.1% accuracy (SD=2.8%), 91.3% sensitivity (SD=3.4%), 98.6% specificity (SD=3.2%), and an F1 score of 94.9% (SD=2.6%). For the four types of birth analysis, RF delivers the best results using all sociodemographic and signal features, achieving 93.5% accuracy (SD=3.0%), mean sensitivity of 95.4% (SD=5.3%), mean specificity of 89.8% (SD=1.8%), and an F1 score of 95.1% (SD=5.6%).

This project successfully distinguishes between different labor types, offering a non-invasive, robust, and low-cost solution. It has the potential to reduce neonatal and maternal mortality, optimize hospital resources, enhance efficiency, improve clinical decision-making, and reduce complications, aligning with several UN Sustainable Development Goals.