Intra-group and inter-group electrocardiograph coding image fusion and classification based on multi-scale group convolution feature fusion network

https://doi.org/10.1016/j.bspc.2021.103374Get rights and content

Highlights

  • A multi-scale group convolutional feature fusion classification model was proposed.

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

  • Weighted sequence fusion was used to realize inter-group and intra-group fusion.

Abstract

Electrocardiograph (ECG) signals exhibit outstanding advantages in the diagnosis of cardiovascular diseases, but their classification is difficult to perform in present studies. The first lead of 12 lead ECG signal has important scientific significance and convenience on the classification of cardiovascular disease. This study proposes a classification model in which the time domain feature and independent component extracted from the first lead ECG signal were transformed into images, and a multi-scale group convolutional feature fusion (MSGCFF) was used to classify seven signal types. First, the time-domain features and principal component features of ECG signals are extracted and thereafter encoded into ECG images using the Gramian angular fields coding method. Next, the depth features of different output layers are extracted by training the ECG image data with the MSGCFF model, and the features are fused by the parallel method. Finally, the fused features were classified using the Softmax classifier. The experimental results were evaluated using the Chinese cardiovascular disease database dataset. The accuracy, precision, sensitivity, and F1 scores of the method were 93.0533%, 81.6354%, 94.8723, and 85.9987%, respectively. This method effectively reconstructs the ECG signal, which not only retain the time dependence, but also utilized the advantages of convolution operation to capture multi-scale features, providing a feasible method for the study of cardiovascular disease classification.

Introduction

Electrocardiogram (ECG) signals are non-stationary and nonlinear. Because ECG signals are reliable, non-invasive, simple, and can provide valuable and rich information [1], they have become the first choice for the detection and diagnosis of cardiovascular and cerebrovascular diseases and can be used as the basis for timely treatment of patients. However, in the current research, ECG is either used as a short-time and stable signal or as an extracted digital feature for analysis, ignoring the signal time domain, spatial domain, and frequency domain features. Therefore, it is impossible to measure the objectivity and integrity of time-varying signal feature extraction [2]. In addition, ECG signals are affected by individual differences and noise in the acquisition process, such that the signals contain a significant amount of irrelevant information [3]. This shows that the current research hotspot is the classification of the data accurately and obtaining more accurate ECG signals.

Convolutional neural networks (CNNs) are considered effective data analysis and processing methods in which special convolution and pooling blocks are used to extract data features, achieve feature reconstruction at the full connection layer, and complete classification with classifiers [4], [5], [6], [7]. In recent years, studies on ECG signal classification using CNN models have been increasing, and some achievements have been realized. Hammad et al. [8] proposed a CNN biometric authentication model and algorithm based on an ECG dataset to address the problem of real-time detection and extract significant features. The experimental results demonstrate that this model is superior to traditional methods. Nguyen et al. [9] used a model combining CNN and Support Vector Machine to predict atrial fibrillation by analyzing the ECG signals. Under 5-fold cross-validation, the average score of F1 was 84.19%. However, this approach only focuses on the statistical characteristics of ECG signals. The rise of one-dimensional (1-D) convolutional networks creates conditions for extracting local features of single leads. A 1-D convolution neural network was proposed for implementation of 17 classes for more than 10 s. The results demonstrate that the model can be expanded to the field of telemedicine and the development of a cloud platform [10]. However, 1-D signal analysis ignores the correlation between the leads in the ECG signal. Acharya et al. [11] designed a deep CNN to check ECG signals at different time periods, and the four groups of Conv-pooling layers constitute the basic architecture of the network. Through experimental verification and analysis, the evaluation index of the model was stable and feasible. However, ECG signals are time series signals, which are highly dependent on the time dimension and have timeliness and uncertainty. Different CNN convolution kernel mappings ignore the coherence and structure of ECG signals and destroy the time correlation of signals, which limits the application of existing CNN models to ECG signal processing. The CNN model focuses more on the extraction of spatial domain and multi-scale features. The Gramian angular fields (GAFs) method is used to map ECG signals to image data, whereas the multi-scale CNN is used to classify ECG images [12]. The advantages of CNN model feature extraction can be fully utilized to obtain a more accurate classification performance. However, this study has not been previously conducted.

In this study, a multi-scale group convolutional feature fusion (MSGCFF) classification model was proposed to classify seven signal types. The optimal fusion ratio of inter-group fusion and intra-group fusion of ECG images is discussed by mapping ECG signals onto ECG color images according to two preprocessing methods, and the classification accuracy is compared with the ECG signal, common CNN, and LeNet-5 neural network. The experimental results prove that it is feasible to transform the time series signals into two-dimensional (2-D) images to realize the classification task, and effectively improve the recognition accuracy of various diseases.

Section snippets

Material and methods

This paper presents a feature fusion and classification method that maps 1-lead ECG signals onto 2-D images. First, the characteristics of ECG signals were extracted: (1) Recurrent neural network (RNN) extracted the time-domain characteristics of signals; (2) Fast independent component correlation (F-ICA) is used to extract the spectral features of the signal, and the dimension reduction of the spectral features is achieved by principal component analysis (PCA). Thereafter, the signal sequences

Experimental design

Two groups of sample data were classified and identified according to the 4-fold cross-validation method. For the two groups of sample data, seven pattern samples were divided into training and testing sets in a proportion of 33%. The mean value obtained by cross-validation was used as the final evaluation index.

The batch size of the MSGCFF model training data was set as 64, number of iterations was set as 40, and learning rate was 1e-3. The number of iterations of the test set was 40, batch

Conclusion

In this study, we propose an MSGCFF model, which is used to realize disease classification. It exhibits good performance in the ECG-encoded image for classification. In particular, the 1-D ECG signal is encoded into two types of 2-D images according to different preprocessing methods, and two groups of features are extracted using a multi-scale convolution kernel (5×53×3, 1×71×3). The best feature fusion method with different weighting ratios (0.1–0.9) is explored using a weighted feature

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.

Acknowledgement

None.

References (45)

  • P. Pławiak et al.

    Novel deep genetic ensemble of classifiers for arrhythmia detection using ECG signals

    Neural Comput. Appl.

    (2020)
  • J. Meyer, K.U. Simmer, K.D.J.A.S. Kammeyer, S.P.i.i.C. on, Multi-channel speech enhancement in a car environment using...
  • Anuradha P et al.

    Noise removal of ECG signals with adaptive filtering

    Mater. Today Proc.

    (2021)
  • R.D. Labati, E. Mu?Oz, V. Piuri, R. Sassi, F.J.P.R.L. Scotti, Deep-ECG: Convolutional Neural Networks for ECG biometric...
  • U.B. Baloglu et al.

    Classification of myocardial infarction with multi-lead ECG signals and deep CNN

    Pattern Recogn. Lett.

    (2019)
  • Y. Wu F. Yang Y. Liu X. Zha S.J.C.p.A.I.C.o.t.I.E.i.M. Yuan, B.S.I.E.i. Medicine, B.S. Conference, A Comparison of 1-D...
  • U.R. Acharya, H. Fujita, O.S. Lih, Y. Hagiwara, J.H. Tan, M.J.I.S. Adam, Automated detection of arrhythmias using...
  • M. Hammad et al.

    A novel two-dimensional ECG feature extraction and classification algorithm based on convolution neural network for human authentication

    Future Gener. Comput. Sys. Int. J. Sci.

    (2019)
  • Q.H. Nguyen et al.

    Stacking segment-based CNN with SVM for recognition of atrial fibrillation from single-lead ECG recordings

    Biomed. Signal Process. Control

    (2021)
  • Z. Yldrm, P. Pawiak, S.T. Ru, U.R.J.C.i.B. Acharya, Medicine, Arrhythmia Detection Using Deep Convolutional Neural...
  • U.R. Acharya, H. Fujita, O.S. Lih, M. Adam, J.H. Tan, C.K.J.K.-B.S. Chua, Automated detection of coronary artery...
  • Z. Wang, T. Oates, Encoding Time Series as Images for Visual Inspection and Classification Using Tiled Convolutional...
  • J.W. Zhang, X. Liu, J.J.I.J.o.A.I.T. Dong, CCDD: AN ENHANCED STANDARD ECG DATABASE WITH ITS MANAGEMENT AND ANNOTATION...
  • R. Reeve, SINUS ARRHYTHMIA, (1960)....
  • Jennifer M. Sims et al.

    Sinus tachycardia

    Nursing

    (1996)
  • M. Chauvin, C.J.P. Brechenmacher, C. Electrophysiology, Atrial Refractory Periods after Atrial Premature Beats in...
  • H. Hashimoto, M. Asano, M.J.J.o.E. Nakashima, Potentiating effects of a ventricular premature beat on the alternation...
  • J. Brugada et al.

    Long-term follow-up of individuals with the electrocardiographic pattern of right bundle-branch block and ST-segment elevation in precordial leads V1 to V3

    Circulation

    (2002)
  • Sergei F. Pravdin et al.

    Overdrive pacing of spiral waves in a model of human ventricular tissue

    Sci. Rep.

    (2020)
  • X, Li, Classification of First Lead Electrocardiogram Based on Convolutional Neural Networks [D] 2019 Tianjin...
  • D. Valencia, D. Orejuela, J. Salazar, J. Valencia, Comparison analysis between rigrsure, sqtwolog, heursure and...
  • A. Ullah, S.U. Rehman, S. Tu, R.M. Mehmood, M.J.S. Ehatisham-Ul-Haq, A Hybrid Deep CNN Model for Abnormal Arrhythmia...
  • Cited by (4)

    • Comparison of the representational ability in individual difference analysis using 2-D time-series image and time-series feature patterns

      2023, Expert Systems with Applications
      Citation Excerpt :

      It can be seen from Table 16 that different ECG preprocessing (conversion of the signal to the image) methods have achieved better classification results on different datasets. Most studies on the conversion of ECG signals to images are based on heartbeat and QRS band segmentation and reconstruction (Huang & Wu, 2020; Mukhopadhyay & Krishnan, 2022; Naz et al., 2021; Wasimuddin et al., 2019), and a few studies consider the integrity of time information (Li & Wang, 2022). Naz et al. (Naz, et al., 2021) used the AlexNet, VGG-16, and Inception-v3 models to extract features for studying ventricular arrhythmias disease on the MIT-BIH dataset.

    View full text