Original Research ArticleSpectral entropy and deep convolutional neural network for ECG beat classification
Introduction
Abnormal heart conditions can cause sudden cardiac death, hence, the early detection of them is vital to identify heart problems. Predictive models developed by emerging technologies in science and engineering can help for accurate diagnosis in medicine. Computer-aided diagnosis (CAD) methods for heartbeat classification is one of the most popular topics in healthcare and bioinformatics to heart-related abnormality detection in electrocardiogram (ECG) signals [1,2].
Analysis of ECG signal provides valuable information about heart conditions of persons for clinicians. It was depicted that the ECG signals of persons with similar heart conditions are almost similar, which makes it possible to predict arrhythmias by analyzing the patterns of ECG signals. On the other side, ECG signals have rapid variations in amplitude, duration, and morphology and these variations make the classification of them as a challenging task for arrhythmia detection. Therefore, developing robust and efficient CAD tools are necessary for early detection of arrhythmias to prevent sudden cardiac death [3,4].
There are several research areas that used ECG signal for different purposes such as diagnosing heart abnormalities [5,6], biometric identification [7], emotion recognition [8], and examining the rhythm of heartbeats. Here, we focus on those used for heart arrhythmia detection.
In some works, higher-order statistics were considered as efficient features for heartbeat diagnosis. The cumulants of the second, third, and fourth orders were considered as features in [9] and then, fuzzy hybrid neural networks have been used for arrhythmia detection. Six types of arrhythmias including the ventricular flatter wave (VF), the left bundle branch block (LBBB), the right bundle branch block (RBBB), premature ventricular ectopic beat (VEB), the atrial premature beat (APB), and the ventricular escape (VE) along with normal ECG signals from MIT-BIH dataset beat were used for performance evaluation. In ref. [10], Hermite coefficients and higher-order statistics construct the feature vector and support vector machine (SVM) classifier was utilized for classification. To evaluate the performance of the introduced method, the authors considered the normal signals and 12 arrhythmia types from MIT-BIH dataset. In ref. [11], after five-level of discrete wavelet transform (DWT) decomposition of the ECG signal, the four sets of higher-order statistics features from the three mid-band components were computed. After feature selection using correlation coefficient, the feed-forward neural network was used for classification of five types of arrhythmias, namely APB, LBBB, paced beats (PB), premature ventricular contractions (PVC), and RBBB and normal heartbeats.
Neural network-based classifiers were used in several works. In ref. [12], DWT with different mother wavelet was used for feature extraction. Then, the performances of the radial basis function neural network (RBFNN), feed-forward network (FFN), and back-propagation neural network (BPN) were compared, where ECG signals, which were chosen from MIT-BIH dataset, were classified into normal or abnormal classes. Deterministic learning was considered in ref. [13] to classify five types of ECG signals from MIT-BIH dataset including normal, LBBB, RBBB, PVC, and PB. Evolvable block-based neural networks (BbNNs), which consist of a two-dimensional (2D) array of modular component, the structure was used in ref. [14] for patient-specific. The Hermite transform coefficients and the RR interval were computed as features from the ECG signals of the MIT-BIH dataset, where five types according to AAMI standard were chosen including normal, ventricular ectopic beats (VEB), supraventricular ectopic beats (SVEB), ventricular fusion (F), and unknown beats (Q).
In some works, deep classifiers were used to heartbeat diagnosis. Convolutional neural network (CNN) classifier is one popular deep classifier in biomedical applications. In ref. [15], ECGs were recorded with a single-lead wearable monitor. A 34-layer CNN was trained to classify 12 heart arrhythmias, sinus rhythm, and noise. Authors in ref. [16] developed a deep CNN to classify 12 rhythm types including 10 arrhythmias as well as sinus rhythm and noise. The used dataset includes 91,232 ECG records obtained by wearable devices from 53,549 patients. The work in ref. [17] considered the VEB and SVEB (or S) beats from the MIT-BIH dataset. The beats were transformed into dual beat coupling matrix according to heartbeat rate and then were given to CNN. SVEB and VEB heartbeats from MIT-BIH dataset was used to obtain the performance. In ref. [18], at first a generic CNN (GCNN) was trained, and then, the GCNN was modified to a dedicated CNN using the fine-tuning technique. This research considered the VEB, SVEB, F, and Q beats from MIT-BIH dataset. In ref. [19], patient-specific ECG classification was proposed, where feature extraction and classification were done using one-dimensional CNN, where VEB and SVEB heartbeats from MIT-BIH dataset was considered. Authors in ref. [20] used the 10-sec segments of ECG signals from the MIT-BIH dataset instead of utilizing one QRS complex and used a CNN for arrhythmia detection. In ref. [21], 44 ECG records from MIT-BIH dataset were considered into five classes according to AAMI standard and a CNN was used to classify them.
In ref. [22], stacked denoising autoencoders (SDAEs) was used for feature extraction from ECG signals. Then, a deep neural network was created by adding a softmax regression layer to classify the ECG signals from MIT-BIH, INCART, and SVDB datasets. A similar method to [22] was employed in ref. [23] for feature extraction and classification in which the breaking-ties and modified version of it were used to choose the informative samples. In addition to MIT-BIH dataset, ECG signals acquired from the wearable device were used to evaluate the method. Authors in ref. [24] proposed to combine a CNN with a denoising autoencoder (DAE) to extract more robust features. They combined the encoder part of the DAE with a fully connected layer of CNN and the resulting classifier is named as convolutional denoising autoencoder (CDAE). The performance of the introduced CDAE was evaluated considering the VEB rhymes from MIT-BIH dataset. In ref. [25], features were extracted based on the temporal and morphological information using the recurrent neural network (RNN) and then global and updatable RNN was introduced to the feature learning and optimization mechanism. Deep belief networks and restricted Boltzmann machine were used in ref. [26] to classify the VEB and SVEB heartbeats from the MIT-BIH dataset.
Some works considered time-frequency transforms for analyzing ECG signals. In ref. [27], after preprocessing by delayed error normalized least mean square (LMS) adaptive filter, DWT was applied for feature extraction from hear-beat variability (HRV) signals. Finally, the ECG signal was classified into normal and abnormal subjects using SVM classifier. In ref. [28], DWT was used to extract morphological features and then, the number of features was reduced by principal component analysis (PCA). Finally, SVM was utilized to classify ECG signals of MIT-BIH dataset into five categories. An improved multiresolution wavelet transform was utilized in ref. [29] to extract the features from QRS complex of ECG signals chosen from MIT-BIH dataset. Neural network and SVM classifier were used to classify the cardiac abnormalities including LBBB, RBBB, and paced beats as well as normal sinus rhythms were classified using. In ref. [30], morphological, statistical, and temporal were computed from DWT of ECG signal. To make a decision, a decision-level fusion of features was considered to classify the heartbeats of the MIT-BIH dataset. A combination of multiple SVMs was presented in ref. [31] that considers the RR intervals and their morphology from ECG signals of MIT-BIT dataset. The final decision was obtained considering the sum, product, and majority rules of the output of each SVM.
In ref. [32], ECG signals were decomposed into intrinsic mode functions (IMFs) using empirical mode decomposition (EMD) and ensemble EMD (EEMD) approaches separately. Coefficient of variation, sample entropy, singular values, and band power of each IMF construct the feature vector. Computed features were given to sequential minimal optimization SVM (SMO-SVM). The MIT-BIH and INCART datasets were considered for performance evaluation, where the aim was to distinguish five types of heartbeat types including normal, PVC, atrial premature contraction (APC), LBBB, and RBBB. The non-stationary and nonlinear analysis, namely improved complete EEMD (ICEEMD), was used in ref. [33] to decompose the EEG signals. The higher-order statistics and sample entropy were fed as features to adaptive boosting ensemble classifier for classifying five types of beats.
As mentioned previously, the diagnosis of ECG heartbeats is a fundamental task to monitor the operation of the heart and can help to prevent sudden cardiac death and early prediction of arrhythmia. It was shown that ECG signals are nonlinear, hence traditional linear signal processing methods cannot capture the frequency content of these signal. Therefore, in this study, we propose to use time-frequency method to analyze the frequency-domain content of heartbeats of the ECG signal at different times. Short-time Fourier transform (STFT) and DWT are the most used time-frequency approaches to analyze the biological signals. On the other side, entropy is an efficient measurement that has been used to diagnose the biological signals [34]. Accordingly, we propose to use entropy in time-frequency manner. To this end, we propose to calculate the spectral entropy resulting in three-dimensional time-frequency transform (TFT) for QRS complexes. It is possible that heartbeats from different classes have the same content in some time-frequency points. To ignore the irrelevant features and increase the separation between TFT of different classes, we use the two-directional two-dimensional PCA (2D2PCA), which performs PCA simultaneously on rows and columns of TFT. Finally, the convolutional neural network (CNN), which is the most popular deep neural network, is used to diagnose the ECG heartbeat. We consider the five arrhythmias from MIT-BIT dataset as well as the normal sinus rhythm. The results demonstrate that the proposed method achieves the proper separation between different types of ECG signals and outperforms the recently introduced method for ECG beat classification.
Following this introduction, Section 2 describes the used dataset in this study and the considered arrhythmias. The proposed method for arrhythmia detection is explained in Section 3. Section 4 provides the results of performance evaluation and finally, concluding remarks are given in Section 5.
Section snippets
ECG signals
The MIT-BIH arrhythmia dataset [35] is considered in this study to assess the performance of the proposed CAD method for heartbeat classification. The 47 individuals contributed to collect this dataset and 48 two-channel (MLII and V5) records with the duration of 24 h were obtained and then 30 min of each record was selected. Then, these continuous signals were passed through a bandpass filter in the range [0.1 100] Hz and after that digital data were obtained by the 360 Hz sampling rate. There
Proposed method
Here, we explain the proposed method for ECG classification in detail. According to Fig. 1, the three main parts of the proposed method are as follows: 1) pre-processing of ECG signal, 2) feature extraction from QRS complex, and 3) classification of time-frequency features.
Experiments and results
Here we explain the experiments carried to assess the performance of the proposed ECG classification. Since ECG signals have autocorrelation, the obtained results with the random splitting of dataset into train and test data may be wrong [48]. Therefore, we consider the hold-out validating instead of k-fold cross-validation scheme since k-fold scheme splits the dataset randomly into k partitions. To perform the hold-out validation, the dataset is partitioned into training and test dataset
Conclusion
A new method based on time-frequency analysis of the QRS complex was proposed in this paper for heart arrhythmia detection. After time-frequency analysis of ECG signals by spectral entropy, the irrelevant time-frequency samples were removed by 2D2PCA and the remaining samples are given to the trained CNN for arrhythmia detection. In this paper, five types of arrhythmias including normal, APB, LBBB, PVC, and RBBB are considered. The performance of the proposed method is obtained for different
References (50)
- et al.
ECG-based heartbeat classification for arrhythmia detection: a survey
Comput Methods Programs Biomed
(2016) - et al.
An automated ECG signal quality assessment method for unsupervised diagnostic systems
Biocybern Biomed Eng
(2018) Application of stacked convolutional and long short-term memory network for accurate identification of CAD ECG signals
Comput Biol Med
(2018)- et al.
Cyclic spectral analysis of electrocardiogram signals based on GARCH model
Biomed Signal Process Control
(2017) Human identification using finger vein and ECG signals
Neurocomputing
(2019)- et al.
An accurate emotion recognition system using ECG and GSR signals and matching pursuit method
Biomed J
(2017) - et al.
Noise-tolerant electrocardiogram beat classification based on higher order statistics of subband components
Artif Intell Med
(2009) - et al.
Efficient wavelet families for ECG classification using neural classifiers
Procedia Comput Sci
(2018) - et al.
ECG beat classification via deterministic learning
Neurocomputing
(2017) - et al.
Patient-specific ECG classification by deeper CNN from generic to dedicated
Neurocomputing
(2018)
Multiresolution wavelet transform based feature extraction and ECG classification to detect cardiac abnormalities
Measurement
ECG classification using three-level fusion of different feature descriptors
Expert Syst Appl
Heartbeat classification fusing temporal and morphological information of ECGs via ensemble of classifiers
Biomed Signal Process Control
Classification of ECG heartbeats using nonlinear decomposition methods and support vector machine
Comput Biol Med
Classification of imbalanced ECG beats using re-sampling techniques and AdaBoost ensemble classifier
Biomed Signal Process Control
Entropies for automated detection of coronary artery disease using ECG signals: a review
Biocybern Biomed Eng
Automated diagnosis of arrhythmia using combination of CNN and LSTM techniques with variable length heart beats
Comput Biol Med
A review of gear fault diagnosis using various condition indicators
Procedia Eng
(2D) 2PCA: two-directional two-dimensional PCA for efficient face representation and recognition
Neurocomputing
Gray-level co-occurrence matrix of Fourier synchro-squeezed transform for epileptic seizure detection
Biocybern Biomed Eng
A hybrid method for MRI brain image classification
Expert Syst Appl
Cardiac arrhythmia classification using the phase space sorted by Poincare sections
Biocybern Biomed Eng
Electrocardiogram signal modeling with adaptive parameter estimation using sequential bayesian methods
IEEE Trans Signal Processing
ECG beat recognition using fuzzy hybrid neural network
IEEE Trans Biomed Eng
Support vector machine-based expert system for reliable heartbeat recognition
IEEE Trans Biomed Eng
Cited by (48)
OptRPC: A novel and optimized recurrence plot-based system for ECG beat classification
2022, Biomedical Signal Processing and ControlCitation Excerpt :In this paper, both datasets are used to better examine our approach In DS1, 16 different labels were grouped into 5 classes: Normal beats (N), Supraventricular ectopic beats (S), Ventricular ectopic beats (V), Fusion beats (F), Unknown beats (Q). In DS2, five annotation labels are chosen in accordance with [41–43] and the classes are as follows: Normal beats (N), Left Bundle Branch Block (LBBB), Right Bundle Branch Block (RBBB), Premature Ventricular Contraction (PVC) and Aberrated atrial Premature Beats (APB). Our proposed scheme consists of four parts: Segmentation, Reconstruction of Phase Space, Recurrence Plot, Classification.
Detection of Atrial Fibrillation from Holter ECG using 1D Convolutional Neural Network after Arrhythmia Extraction
2024, Advanced Biomedical EngineeringHeartbeat Classification Based on PCA and CNN
2024, IFMBE ProceedingsHybrid generative model for grading the severity of diabetic retinopathy images
2024, Computer Methods in Biomechanics and Biomedical Engineering: Imaging and Visualization