Enhancing classification of preterm-term birth using continuous wavelet transform and entropy-based methods of electrohysterogram signals

Abstract

Introduction: Despite vast research, premature birth's electrophysiological mechanisms are not fully understood. Prediction of preterm birth contributes to child survival by providing timely and skilled care to both mother and child. Electrohysterography is an affordable, noninvasive technique that has been highly sensitive in diagnosing preterm labor. This study aimed to choose the more appropriate combination of characteristics, such as electrode channel and bandwidth, as well as those linear, time-frequency, and nonlinear features of the electrohysterogram (EHG) for predicting preterm birth using classifiers.

Methods: We analyzed two open-access datasets of 30 minutes of EHG obtained in regular checkups of women around 31 weeks of pregnancy who experienced premature labor (P) and term labor (T). The current approach filtered the raw EHGs in three relevant frequency subbands (0.3-1 Hz, 1-2 Hz, and 2-3Hz). The EHG time series were then segmented to create 120-second windows, from which individual characteristics were calculated. The linear, time-frequency, and nonlinear indices of EHG of each combination (channel-filter) were fed to different classifiers using feature selection techniques.

Results: The best performance, i.e., 88.52% accuracy, 83.83% sensitivity, and 93.22% specificity, was obtained in the 2-3 Hz bands using Medium Frequency, Continuous Wavelet Transform (CWT), and entropy-based indices. Interestingly, CWT features were significantly different in all filter-channel combinations. The proposed study uses small samples of EHG signals to diagnose preterm birth accurately, showing their potential application in the clinical environment.

Discussion: Our results suggest that CWT and novel entropy-based features of EHG could be suitable descriptors for analyzing and understanding the complex nature of preterm labor mechanisms.

Keywords: electrohysterography; entropy; machine learning; preterm labor; time-frequency analysis; uterine electromyogram.

Figure 1

Block diagram of the proposed methodology for P and T classification.

Figure 2

Placement of electrodes in the maternal abdomen (modified from 7).

Figure 3

Selection of features for each classifier. S1, S2, and S3 represent the bipolar channels acquired from the public database, while F1, F2, and F3, represent the frequency sub-bands compared in this study. Colored frames (gray, yellow, and orange) indicate that the feature was selected as an optimal classification parameter and used to train the prediction model. The classifiers that achieved the highest classification accuracy (< 85%) in testing are presented in yellow and orange, indicating the type of classifier used. *,**,***,**** are used to describe significant differences between P and T groups in the Mann-Whitney test, with p-values lower than 0.05, 0.01, 0.001, and 0.0001, respectively.

Figure 4

Energy values for the channel 2 in the three different subbands. The three panels can corroborate the difference between the P and T groups, where the T group presents higher values than P. (A) Values for the S2F1 combination. (B) Shows the values of Energy for S2F2 and (C) the Energy values for S2F3. **** represents p-value lower than 0.0001.

Figure 5

Representative spectrograms of EHG signals for the term (T) and preterm (P) groups. (A) shows the spectrogram retrieved from the participant no. 10, from the preterm group (P, recorded at 30 weeks of gestation and triggered labor at 32 weeks of gestation), while (B) shows the spectrogram retrieved from participant no. 9, from the term group (T, recorded at 30 weeks of gestation and triggered labor on 39 weeks of gestation). Below each spectrogram, the EHG signal can be visualized. Both spectrograms are derived from the S2 signal.