Predicting preterm labour is a serious and unresolved problem. In this
master thesis I focused on resolving the problem of separation and classifi-
cation of groups of uterine electromyographic recordings with the term and
pre-term birth, which were recorded before 26th week and after 26th week of
gestation. Until now, studies have been made where they separated the two
groups in the frequency range with a lower limit of 0.08 Hz to 0.3 Hz and
highest limit of 3 Hz or 4 Hz. In this master thesis, I focused on the area
between 0.08 Hz and 5 Hz, since these areas are not well researched. I used
the sample entropy as one of the currently most promising techniques and
international reference database TPEHG DB of uterine electromyographic
recordings. For the classification of recordings I tested the K nearest neighbors
(K-NN) classifier, linear and quadratic discriminant analysis (LDA,
QDA) classifiers, Naive Bayes, support vector machine method and decision
trees. To assess the degree of separation of groups of recordings I used a
statistical analysis of variance (ANOVA). Because of the uneven distribution
of the number of recordings that were preterm and term, I used Synthetic
Minority Over-sampling Technique (SMOTE), in order to ensure more real
results. The result of the master’s thesis is that it was confirmed that the
classification between preterm and term births on the recordings that were
recorded early the best area was from 1 Hz - 5 Hz. We have also found that
by increasing of the frequency or by widening of the frequency area, preterm
recordings become less regular and less predictable; term recordings on the
contrary, become more regular and more predictable. The achieved results of
classification of recordings with term and preterm delivery using frequency
band from 1 Hz and 2.2 Hz, and sample entropy features only, (Sensitivity
88.8 %, Specificity 81.8 %, Classification accuracy 85.3 %) are quite comparable
to achieved results of classification of other existing studies (Sensitivity
89 %, Specificity 91 %, Classification accuracy 90 %) which, besides sample
entropy features, also involved other signal features, signal spectra features,
as well as additional clinical information about pregnancies.
|