Unsupervised Classification of Uterine Contractions Recorded Using Electrohysterography

Abstract

Pregnancy still poses health risks that are not attended to by current clinical practice motorization procedures. Electrohysterography (EHG) record signals are analyzed in the course of this thesis as a contribution and effort to evaluate their suitability for pregnancy monitoring. The presented work is a contributes with an unsupervised classification solution for uterine contractile segments to FCT’s Uterine Explorer (UEX) project, which explores analysis procedures for EHG records. In a first part, applied processing procedures are presented and a brief exploration of the best practices for these. The procedures include those to elevate the representation of uterine events relevant characteristics, ease further computation requirements, extraction of contractile segments and spectral estimation. More detail is put into the study of which characteristics should be chosen to represent uterine events in the classification process and feature selection methods. To such end, it is presented the application of a principal component analysis (PCA) to three sets: interpolated contractile events, contractions power spectral densities, and to a number of computed features that attempt evidencing time, spectral and non-linear characteristics usually used in EHG related studies. Subsequently, a wrapper model approach is presented as a mean to optimize the feature set through cyclically attempting the removal and re-addition of features based on clustering results. This approach takes advantage of the fact that one class is known beforehand to use its classification accuracy as the criteria that defines whether the modification made to the feature set was ominous. Furthermore, this work also includes the implementation of a visualization tool that allows inspecting the effect of each processing procedure, the uterine events detected by different methods and clusters they were associated to by the final iteration of the wrapper model.