An automatic classification approach for preterm delivery detection based on deep learning

Abstract

Recently, tocography (TOCO) and electrohysterogram (EHG) signals are real-time and non-invasive technology that has been applied to detect preterm delivery. This paper proposes a novel deep learning (DL) framework for automatically detecting preterm delivery based on EHG data. Initially, the recorded EHG signal, which corresponds to the electrical activity of the uterine muscles, is pre-processed using a 4th-order Butterworth filter and 6th-order Daubechies wavelet transform for filtering and denoising. Then, the pre-processed signal is employed for the feature extraction stage, mainly to extract the significant features specifically, median frequency, peak frequency, sample entropy, spectral entropy, power spectral density (PSD), fuzzy entropy, root mean square (RMS), and Shannon entropy. Next, the classification is performed using a two-step scheme called a circle-optimized bidirectional long short-term gated recurrent model (CoBi-LSTGR) to learn the extracted features. Here, the extracted features are initially fed to the bidirectional short-term memory (Bi-LSTM) and then provided to the bidirectional gated recurrent unit (Bi-GRU) model for more precise learning. At last, the learned features are used to classify the EHG signal into term and preterm. Subsequently, the hyperparameter tuning in both the learning models is done using the circle-inspired optimization algorithm (CiOA). The proposed framework is implemented in the Python platform and evaluates various performance metrics to validate the competence of a proposed scheme by employing the Term-Preterm EHG (TPEHG) database. The accuracy attained by a proposed for channels I, II, and III are found to be 99.09%, 99.01% and 99.89%, respectively. The experimental outcomes show that the proposed framework performs better against existing classifiers.