Single-scale convolution wavelet feature optimization classification model based on electrocardiogram coded image

https://doi.org/10.1016/j.bspc.2022.104202Get rights and content

Highlights

  • A single-scale convolution wavelet feature optimization model was proposed.

  • 1-D ECG signal was encoded into 2-D coded images.

  • It is better to distinguish diseases with similar distribution characteristics.

  • SSCWFO model are improved to 91.12%, 92.29%, 92.55%, and 92.13%, respectively.

Abstract

Premature beats are a cardiovascular disease, which can lead to complications of other diseases. An electrocardiograph (ECG) is the main tool for detecting premature beats, but accurate detection of premature beats still faces challenges. This work designs a single-scale convolution wavelet feature optimization (SSCWFO) classification model that is based on a single-scale convolutional neural network (SSCNN) model and optimized by Linear Discriminant Analysis (LDA) algorithm. First, the wavelet coefficient features are extracted from 12-lead ECG signals using Daubechies5, and then the 5th level detail coefficient features are transformed into images using Gramian Angular Difference Fields (GADFs). Thereafter the single-scale features are extracted by the SSCNN model. Next, the LDA algorithm was used to maximum inter-class distance and the minimum intra-class distance of each category. Finally, the Gaussian naive Bayes is used to classify-three types of signals. The results after LDA optimization show that the accuracy, precision, sensitivity, F1-score, and area under the ROC curve (AUC) of the SSCWFO model are improved to 91.12%, 92.29%, 92.55%, 92.13%, and 0.9197, respectively. Also, the network architectures of Resnet34, Resnet18, and LeNet 5 are used in this work for a comparison; their accuracy is 75.93%, 76.88%, and 68.48%, respectively. This shows that the method can effectively distinguish normal signals and two types of premature beat diseases and is helpful to the effective classification of diseases of premature beats.

Introduction

Premature beats, including ventricular premature beats (VPB) and atrial premature beats (APB), are considered a common clinical ventricular disease [1]. And the VPB and APB refer to the lesion of the ventricle and atria, respectively [2]. The distinguished method for them is with aid of the electrocardiograph (ECG) signal that is non-invasive, accurate, and portable [3].

ECG signal is a time-varying signal, which is with the advantages of non-stationary, nonlinear, strong randomness, and continuous change in time [4]. The spatial structure is attached to the multi-lead ECG signal, which makes each lead signal depend on and interact with each other in the spatial dimension. However, the nonlinear and time-varying of ECG signals are easy to be interfered with the time factor interference or signal linearization, leading to an obstruction to the analysis of the VPB and APB. In the current signal analysis, time-varying signals are simply regarded as numerical operations, ignoring the accumulation process of time factors of biological signals, which may be missed important information. As is well-known, the process of information transmission of biological life activities has a certain dependence and continuous correlation on time. Consequently, the time factor should be considered in the research and analysis of ECG signals. The change of time–frequency characteristics of ECG signals is closely related to time factors [5]. Wavelet transform (WT) is used to achieve time–frequency localization analysis of ECG signals, which can focus on any detailed characteristics of ECG signals [6], [7].

Feature extraction is a characterization process in which a group of data is transformed (linear or nonlinear) to extract representative features; it generally refers to the invariance of similar samples and the specificity of different samples, and the extracted features exhibit obvious characterization significance [8], [9]. The frequency-domain feature is characterized as frequency amplitude in two-dimensional (2-D) space, which represents the relationship between ECG signal frequency and amplitude [10]. There are differences in the angle and selection of feature extraction in different models, such as the WT and deep neural network (DNN) model. WT is an optimization algorithm for the time–frequency transformation of ECG signals based on a set of basic functions, which can not only obtain frequency characteristics but also accurately locate the time factor changes [11]. Li et al. [12] used multi-scale wavelet transformation to achieve the detection of characteristic waves; results show that the accuracy of the QRS detection rate of this method is 99.8 %. Kumar et al. [11] decomposed ECG signals and calculated the classification accuracy step by step on the features of decomposition using a flexible wavelet; they also pointed out that this method could be used for large-scale cardiac screening, and the classification accuracy of the Morlet wavelet kernel was 99.60 %. The DNN can automatically learn and characterize features by constantly adjusting and optimizing training parameters through neural feedback. Convolutional neural network (CNN), as a reference network for image feature extraction and classification model, obtains multi-scale features through convolution operation and reduces operation parameters by sharing weights [13], [14], [15], [16]. These indicate that it is a feasible and interesting method to convert time-varying signals into images. However, the current research is still in the exploratory stage.

In this work, the nonlinear characteristics and time factor characteristics of ECG signals were retained using the WT for extracting the time–frequency features of the VPB, APB, and normal signals. Then, the time–frequency features were further encoded into ECG images according to the time series signal coding technology and effectively classify them using the multi-channel classification model based on CNN. That is, the ECG signal was decomposed by WT and mapped to ECG color images. The ECG color images were classified by the SSCWFO model as well as the ResNet18, ResNet34, and LeNet 5 for comparison; finally, the optimal algorithm for LDA was discussed. The results show that restructuring the wavelet detail coefficients to images significantly improves the classification performance of the two types of premature beats and normal signals.

Section snippets

Material and methods

In this work, we provided a preprocessing method for ECG signal image reconstruction based on wavelet coefficients that decompose and obtain the wavelet coefficient matrix from the original records using WT to reconstruct into a 2-dimensional (2-D) image using the Gramian Angular Difference Fields (GADFs) algorithm [17]. First, Daubechies 5 (db5) wavelet transform, as a feature selector, decomposes the multidimensional ECG signal into time-varying frequency bands, which provides local

Experimental basis

The experiment uses 10-fold cross-validation to complete the classification task. The loss function selects the multi-classification cross-entropy function [24]. Experimental data were divided into the training set and test set in 70 % and 30 % proportions. Each classification model was trained and tested using Python 3.7. The training and testing parameter setting of the SSCWFO was set as follows: batch size = 100, and the learning rate and iteration times of the classification model are 0.001

Conclusion

This work developed a SSCWFO model that is applied to complete three types of image classification. The results show that the LDA algorithm can effectively distinguish two kinds of premature beats and normal signals by maximizing the inter-class distance and minimizing the intra-class distance. On the one hand, each classification method effectively captures the critical time–frequency features of TEI images. On the other hand, the wavelet transform retains the nonlinear characteristics and

CRediT authorship contribution statement

Jingjing Li: Conceptualization, Methodology, Project administration, Writing – original draft, Writing – review & editing, Data curation, Formal analysis. Qiang Wang: Formal analysis, Investigation, Software, Supervision, Funding acquisition, Resources, Validation.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgements

None.

References (28)

  • S.Z. Mahmoodabadi, A. Ahmadian, M. Abolhasani, M. Eslami, J.H. Bidgoli, ECG Feature Extraction Based on Multiresolution...
  • J.D. Peshave, R. Shastri, Feature extraction of ECG signal, International Conference on Communications & Signal...
  • V. Seena, J. Yomas, A review on feature extraction and denoising of ECG signal using wavelet transform, International...
  • C. Li, C. Zheng, C.J.I.T.o.B.E. Tai, Detection of ECG characteristic points using wavelet transforms, 42 (2002)...
  • Cited by (0)

    View full text