Deep Convolutional Generative Adversarial Network with LSTM for ECG Denoising

Wang, Huidong; Ma, Yurun; Zhang, Aihua; Lin, Dongmei; Qi, Yusheng; Li, Jiaqi

doi:https://doi.org/10.1155/2023/6737102

Computational and Mathematical Methods in Medicine

On this page

Abstract Introduction Materials and Methods Results and Discussion Conclusions Data Availability Conflicts of Interest Acknowledgments References Copyright Related Articles

Research Article | Open Access

Volume 2023 | Article ID 6737102 | https://doi.org/10.1155/2023/6737102

Deep Convolutional Generative Adversarial Network with LSTM for ECG Denoising

Huidong Wang,^1,2,3Yurun Ma ,^1,2,3Aihua Zhang,^1,2,3Dongmei Lin,^1,2,3Yusheng Qi,^1,2,3and Jiaqi Li^1,2,3

Academic Editor: Nagarajan DeivanayagamPillai

Received17 May 2022

Revised23 Nov 2022

Accepted19 Jan 2023

Published10 Feb 2023

Abstract

The electrocardiogram (ECG), as an essential basis for the diagnosis of cardiovascular diseases, is usually disturbed by various noise. To obtain accurate human physiological information from ECG, the denoising and reconstruction of ECG are critical. In this paper, we proposed an ECG denoising method referred to as LSTM-DCGAN which is based on an improved generative adversarial network (GAN). The overall network structure is composed of multiple layers of convolutional networks. Furthermore, the convolutional features can be connected to their time series order dependence by adding LSTM layers after each convolutional layer. To verify the effectiveness and the denoising performance of the improved network structure, we test the proposed algorithm on the famous MIT-BIH Arrhythmia Database with different levels of noise from the MIT-BIH Noise Stress Test Database. Experimental results show that our method can remove the single noise and the mixed noise while retaining the complete ECG information. For the mixed noise removal, the average SNR_imp, RMSE, and PRD are 19.254 dB, 0.028, and 10.350, respectively. Compared with the state-of-the-art methods, DCGAN, and the LSTM-GAN methods, our method obtains the higher SNR_imp and the lower RMSE and PRD scores. These results suggest that the proposed LSTM-DCGAN approach has a significant advantage for ECG processing and application in complex scenes.

1. Introduction

Cardiovascular diseases pose a serious threat to human health, especially due to the sudden and intermittent nature of heart disease; therefore, routine monitoring is the best strategy to diagnose these diseases. As a “gold standard” for assessing heart conditions, ECG has been widely used in clinical diagnosis and daily personal monitoring. However, the captured ECG signals are usually interfered by the noise from the equipment, environment, and human movement, which may lead to a failed diagnosis. Therefore, it is crucial to remove the interference from the recorded ECG and reconstruct a clean ECG.

Generally speaking, the noise affecting the ECG signal mainly includes the electrode motion (EM), the muscle artifact (MA), and the baseline wander (BW). Many scholars have conducted in-depth research on removing noise from ECG signals and proposed some efficient methods. For example, filtering [1–5], wavelet [6], and empirical mode decomposition (EMD) [7] are typical ECG denoising methods. Filtering methods include finite impulse response (FIR) filter [1], infinite impulse response (IIR) filter [2], adaptive filter [3], Kalman filter [4], and Wiener filter [5]. Among them, the FIR and IIR filters are not applicable when the frequency domains of the ECG and noise signals overlap. The adaptive filter requires the reference signal, and the Wiener filter is difficult to remove the complex noise. For the wavelet methods, the choices of wavelet function and threshold function have a great influence on the denoising performance. To reduce the end point effect and modal aliasing in the traditional EMD algorithms, the ensemble empirical mode decomposition (EEMD) [8] and variational mode decomposition (VMD) [9] methods are published. However, it is necessary to artificially add Gaussian white noise in order to use the EEMD approach, and the parameters of the VMD technique are sensitive to varied inputs. There are also several EMD paired with wavelet algorithms for ECG denoising algorithms, but most of them are only useful for certain types of noise reduction [10–12]. In recent years, some researchers have employed machine learning techniques for ECG denoising. Their works can be roughly classified into two categories: (1) improving existing algorithms by machine learning methods for parameter optimization [13, 14] and (2) building the ECG denoising model to decompose and reconstruct the ECG using a neural network [15–19]. To reduce baseline drift, Sun et al. proposed combining error backpropagation neural network and VMD technique [13]. Wang et al. proposed an ECG signal denoising algorithm based on a supervised deep factor analysis model [14]. Xiong et al. used the denoise autoencoder (DAE) for the denoising and compression of ECG signals [15]. Chiang et al. completed the denoising and compression of ECG based on the fully convolutional network denoising autoencoder (FCN-DAE). The fully convolutional network is applied to preserve the detailed information of ECG as much as possible [16]. Dasan and Panneerselvam accomplished ECG compression and reconstruction by adding a long short-term memory (LSTM) network to the encoder tail of a convolutional denoising autoencoder [17]. Wang et al. created the ECG denoising algorithm based on GAN, in which the new loss function and deep neural networks (DNN) were used [18]. However, using DNN is more likely to cause QRS wave group distortion [16]. Therefore, some researchers achieved the denoising model using the residual network structure in GAN [19]. Additionally, the GAN with a CNN-based discriminator and an LSTM-based generator is used to generate ECG [20]. Although the convolutional neural network (CNN) does not cut the waveform amplitude, it reconstructs the denoised signal from the convolutional features; each convolution is independent and lacks mutual information. Because of the significant individual differences, the large pathological differences, and the complex behavioral environments, it is still difficult to remove as much noise as possible while keeping the ECG information intact. Up to now, researchers are still working on developing the ECG denoising approaches which can be integrated into the actual monitoring equipment.

In consideration of both the global information and the local time domain characteristics of ECG, we design the LSTM-DCGAN for ECG denoising. For the local feature capturing, the CNN with a smaller receptive field is used. For the global information reserving, we adopt the GAN structure to ensure that the denoised ECG signal achieves the same data distribution as the clean ECG signal. At the same time, the GAN is improved by adding the LSTM layer after each convolutional layer, which can correlate its sequential dependencies on time series and strengthen its global information capability. The following text will introduce the proposed ECG denoising network from the aspects of methods, experiments, results, and discussion in detail.

2. Materials and Methods

2.1. Denoising Model Based on GAN

The GAN is a structure that estimates a generative model through an adversarial process and is composed of a generator () and a discriminator () [21]. GAN trains two models: the generator generates the observed data by assigning some noise variable , and the discriminator distinguishes between the true data and the generated data . The optimization goal of the generator is to maximize such that the discriminator incorrectly discriminates the generated data. The optimization goal of the discriminator is to maximize as well as minimize to improve its classification performance. Generator and discriminator both improve their performance through adversarial training. This adversarial process is described as a min-max game with the value function shown in Equation (1), until the generator and discriminator have sufficient capacity and the optimization of both reaches a critical state so the generator reproduces the real data distribution. where represents the distribution of real data samples , and represents the distribution of the input variable of the generator, and represents the mathematical expectation about the distribution specified in the subscript.

The noisy ECG signal can be described by where is the ECG signal with noise, is the clean ECG signal, and is noise.

As shown in Figure 1, in the adversarial denoising model, the generator takes the noisy ECG signal as input and outputs the denoised ECG signal . The input of the discriminator contains the denoised ECG signal or the original clean ECG signal , and the discriminator outputs probabilistically by judging whether the input originates from the denoised signal or the clean signal . The ECG denoising problem can be described as obtaining from the noisy ECG signal . where is short for “as close as possible” and is the denoised signal.

The discriminator continuously improves its ability by passing the loss backward to minimize the discriminator . The optimization goal of the generator is to maximize the discriminator by updating the network parameters using loss and loss together, so that the generated denoised signal is as close as possible to the original signal . Both the generator and the discriminator are trained through a continuous game, until the optimization of both the generator and the discriminator reaches a critical state. The denoised signal is almost the same as the clean signal , so the generator has denoising capability.

Because the denoising performance of GAN is determined by the structures of the generator and discriminator, we propose the improved GAN for ECG denoising whose generator structure and discriminator structure are, respectively, described in 2.2 and 2.3.

2.2. Generator and Value Function

It is generally believed that MA and EM noise affect the detailed waveform of ECG. The BW noise affects the overall trend of the ECG signal, which is reflected in the drift of the baseline. Thus, we design the GAN by integrating the deep convolutional generative adversarial networks (DCGAN) [22] and the LSTM [23]. On one hand, we apply the DCGAN replacing the upsampling steps in the original GAN with the fractional-stride convolution to capture the local features and achieve feature recombination for the ECG details. On the other hand, the LSTM controlled by forget gates, input gates, and output gates is used to preserve the global time domain information of ECG. In this way, we design an iterative generator structure consisting of convolutional layers and LSTM layers, which is called CNN-LSTM structure and shown in Figure 2(a). To verify the effectiveness of the designed structure, in 3.1.1, we compare the denoising performance of the CNN-LSTM with the convolutional structure shown in Figure 2(b) and the LSTM structure shown in Figure 2(c).

(a) CNN-LSTM structure

(b) CNN structure

(c) LSTM structure

As shown in Figure 3, the generator consists of an input layer, a hidden layer based on the CNN-LSTM-Block structure, and an output layer. The input layer consists of a fully connected (FC) layer of size , a batch normalization layer, and a LeakyReLU activation layer. ReLU sets all negative values to zero, and the gradient of neurons is always zero. For the training of GAN, ReLU makes the training fragile and a large number of neurons die. Instead, LeakyReLU assigns a nonzero fixed slope to all negative values, ensuring that the gradient can flow through the entire architecture. The LeakyReLU activation function is shown in Equation (4). The slope of PReLU is a nonfixed learnable parameter. The hidden layer consists of 5 CNN-LSTM-Blocks with 128-64-40-20-40 nodes. Each CNN-LSTM-Block consists of one deconvolution layer (filter and ) and one LSTM layer. Batch normalization and LeakyReLU layers are added after each deconvolution layer. After each LSTM layer, a dropout layer with a rate of 0.5 and tanh activation function is added. Tanh activation function is shown in Equation (5). In the output layer, a deconvolution layer of size is used to recover the data length of .

The optimization objective of the generator is to minimize , and the value function is shown in Equation (6). In addition, we focus on the difference between the global and the local of the model by adding distance function () and the maximum local difference (), where and are defined in Equations (7) and (8). where represents the distribution of the noise signal (), represents the number of samples, represents the denoised signal at sample point , and represents the clean signal at sample point , and and are the weighting coefficients of and , which are experimentally set to 0.7 and 0.2, respectively.

2.3. Discriminator and Value Function

As shown in Figure 4, the discriminator consists of an LSTM layer (), a dropout layer (), a convolutional layer (, filter , and ), a max pooling layer ( and ), and a softmax layer.

The optimization goal of the discriminator is to maximize the probability of assigning the correct label to both real examples and generated signals. The value function is shown in

2.4. Experimental Setup

2.4.1. Performance Evaluation Criteria

The improvement of SNR (SNR_imp), the root mean square error (RMSE), and the percentage root mean square difference (PRD) are used to evaluate the effectiveness of the denoising model. Let denotes the clean ECG signal, denotes the noisy ECG signal, and denotes the denoised ECG signal. is the length of the proposed signal in one computing cycle.

SNR_imp characterizes the SNR difference between the denoised signal and the original clean signal, which is defined as follows: where and are, respectively, calculated by Equation (11) and Equation (12). From the definition of SNR_imp, a higher value of SNR_imp corresponds to a better performance.

RMSE denoted in Equation (13) is used to measure the similarity between the denoised signal and the original clean signal. A smaller value of RMSE corresponds to a higher similarity between the denoised signal and the clean signal.

PRD, denoted in Equation (14), is used to measure the error between the reconstructed signal and the target signal. A smaller value of PRD corresponds to a smaller reconstruction loss of the denoising method.

2.4.2. Data Set Experimental Samples

The MIT-BIH Arrhythmia Database [24] and the MIT-BIH Noise Stress Test Database [25] are used to obtain the clean ECG signal data and the noise data, respectively. The MIT-BIH Arrhythmia Database is one of the widely acknowledged standard ECG databases in which a total of 48 records, with 2 channels per record, from 47 subjects were gathered continuously over the course of 24 hours. All ECG signals were digitized at 360 Hz. The MIT-BIH Noise Stress Test Database includes typical noise in 12 half-hour of ECG recordings and 3 half-hour of Holter recordings. Noise recordings are collected from healthy volunteers using a standard ECG acquisition instrument, including BW, MA, and EM. We use the data from the MIT-BIH Arrhythmia Database as clean signals and add noise from the MIT-BIH Noise Stress Test Database to create noisy signals.

(1) Data Set I. The experiment data are obtained from 15 records numbered 100, 103, 106, 109, 115, 116, 123, 202, 205, 209, 220, 223, 230, 231, and 234 in the MIT-BIH Arrhythmia Database. We use the modified lead II (ML II) data for each record because lead II is widely used in portable wearable devices. All signals are segmented according to the length of 1024 samples per fragment, and a total of 8254 fragments are acquired. Data set I is used for model comparison, noise reduction of different noise types, method comparison, and individual variability experiments.

(2) Data Set II. We also use lead V1 data to verify the denoising performance of the model for other leads. Due to the absence of V1 leads in records 100, 103, and 123, experiment data are obtained from records 106, 109, 115, 116, 202, 205, 209, 220, 223, 230, 231, and 234. All signals are segmented into fragments of length 1024 samples and produced 6381 fragments. The studies using new leads are conducted using data set II.

We choose -1 dB, 3 dB, and 7 dB SNR noise to pollute the original clean ECG signal to test the denoising ability of the model at different SNR levels. Before being utilized as input to the model, all signals are normalized and the data amplitude ranges from 0 to 1. We divide the data sets into the train set (95%) and the test set (5%). The following is the normalizing formula: where and are the minimum and maximum values of the sample points.

2.4.3. The Computing Platform

The experimental platform is a Windows environment server with hardware: Intel core i9-9900k CPU at 3.6 GHz with 64 GB of RAM and NVIDIA GeForce RTX2080Ti for computational acceleration. The generative adversarial network is built on Python 3.6 and TensorFlow 2.3. Both the generator and discriminator use the Adam optimizer with an initial learning rate of 0.0001. The batch size is 64.

3. Results and Discussion

3.1. Experimental Results

3.1.1. Model Comparison

The LSTM is used to correlate the temporal sequence of convolutional features to overcome the limitation of weak mutual information in the small receptive field. We compare our method with DCGAN and LSTM-GAN. These three structures are shown in Figure 2. In addition, three activation functions (ReLU, PReLU, and LeakyReLU) are set to confirm the influence of the activation function on the proposed model.

The clean ECG signals for the experiments are from data set I and the noise data from the MIT-BIH Noise Stress Test Database. The noisy ECG signals are formed by adding mixed noise (MA+BW+EM) with different SNR (-1 dB, 3 dB, and 7 dB) to clean ECG signals. The SNR_imp, RMSE, and PRD indicators are used to evaluate the performance of the model. The results of the model structure comparison are shown in Figure 5. The experimental results of the comparison of the three different activation functions in the proposed model are shown in Table 1.

(a) The SNRimp of DCGAN, LSTM-GAN, and LSTM-DCGAN

(b) The RMSE of DCGAN, LSTM-GAN, and LSTM-DCGAN

(c) The PRD of DCGAN, LSTM-GAN, and LSTM-DCGAN

It can be seen from Figure 5 that our proposed convolution combined with LSTM adversarial method has higher SNR_imp and lower RMSE and PRD, regardless of the level of noise pollution. In addition, it can be seen from Table 1 that the model with LeakyReLU as the activation function outperforms the other two models.

3.1.2. Noise Reduction of Different Noise Types

To measure the denoising ability of the model for different noise, we add MA, BW, EM MA+BW, MA+EM, BW+EM, and MA+BW+EM with different SNR levels (-1 dB, 3 dB, and 7 dB) to the clean ECG signals in data set I.

Table 2 shows our method’s average SNR_imp, RMSE, and PRD scores for different kinds of noise removal at -1 dB, 3 dB, and 7 dB. For single noise removal, the SNR_imp of our method reaches 19.873 dB, and for mixed noise removal, it reaches 19.254 dB. The results in Table 2 show that our method can remove not only single noise but also multiple mixed noise. Because EM noise can imitate the appearance of ectopic beats, the results of our model for removing EM noise are not as excellent as other noise.

Figure 6 shows the results of our model for removing EM, BW, and MA. In each subplot, the first row represents the noisy ECG signal. The second row is the comparison between the clean signal (blue) and the denoised signal (red).

(a) The denoising results of ECG with MA noise

(b) The denoising results of ECG with BW noise

(c) The denoising results of ECG with EM noise

Figures 7(a)–7(d) represent the results of our method for removing four kinds of mixed noise, MA+BW, MA+EM, BW+EM, and MA+BW+EM. In each subplot, the first row represents the ECG signal with mixed noise, and the second row compares the denoised signal (red) and the original clean signal (blue). The overall denoised ECG signal is close to the original clean ECG signal, and the experimental results show that our method can also effectively reduce the various mixed noise contained in the ECG signal.

(a) The denoising results of ECG with MA+BW noise

(b) The denoising results of ECG with MA+EM noise

(c) The denoising results of ECG with BW+EM noise

(d) The denoising results of ECG with MA+BW+EM noise

3.1.3. Method Comparison

We compare the proposed method with the state-of-the-art method GAN and FCN-DAE. In GAN [18], the structure of the generator network has the number of nodes 310-250-250-310, and the number of nodes of the discriminator network structure is 310-150-150-1. In the generator and discriminator, a layer of tanh activation function is added after each hidden layer, and the activation function of the output layer is a sigmoid function. In FCN-DAE [16], it includes an encoder () composed of convolutional layers and a decoder () composed of deconvolution layers, and the convolutional kernel size of both encoder and decoder is 16. All methods used clean ECG signal from data set I, and the test set and train set remain consistent. Because the daily collected ECG signal often contains a variety of noise, noisy ECG signals are formed by adding mixed noise (MA+BW+EM). SNR_imp, RMSE, and PRD are used to measure the performance of all methods.

Figures 8(a)–8(c) show the average SNR_imp, RMSE, and PRD scores of the three denoising models. The SNR characterizes the noise level of the signal; a higher value of the SNR corresponds to less noise in the signal. As can be seen from Figure 8(a), for each denoising method, the SNR_imp decreases as the SNR increases. This is consistent with our expectation that as the noise decreases, the degree of noise signal removal also decreases. The RMSE and PRD represent the closeness between the denoised signal and the original clean signal. The lower values of RMSE and PRD correspond to less distortion of the denoised signal. When the SNR is -1 dB, the SNR_imp scores of the three methods are 15.870 dB, 17.569 dB, and 19.254 dB; the RMSE scores are 0.050, 0.043, and 0.035; the PRD scores are 18.77, 15.32, and 12.91. Regardless of the level of SNR, the SNR_imp of our method is higher than that of GAN and FCN-DAE, and the RMSE and PRD of our method are lower than the other two methods, which means the denoising performance of our method is better than the other two methods.

(a) The average SNRimp of GAN, FCN-DAE, and LSTM-DCGAN

(b) The average RMSE of GAN, FCN-DAE, and LSTM-DCGAN

(c) The average PRD of GAN, FCN-DAE, and LSTM-DCGAN

In Section 3.1.1 and Section 3.1.3, we compare the SNR, RMSE, and PRD indicators of the five methods, and the indicators show that our method outperforms several other methods. In Figure 9, we plot the denoising waveforms of the five methods to observe the denoising of the signal. From Figure 9, it can be seen that GAN has a significant cut on the R-wave, which makes the amplitude of the R-wave significantly lower than that of the original clean signal, and the overall waveform has small spurs. The DCGAN can effectively remove the noise from the signal, and the overall waveform is smooth. The R-wave also has a high degree of coincidence with the clean signal but does not match the original waveform in some details. Although LSTM-GAN can reduce very small noise, it is not effective in reducing large amplitude noise. Using FCN-DAE for denoising, the coincidence with the original clean signal is better than the above three methods, but the waveform is not smooth enough. Among the five methods, our method possesses a higher coincidence with the original clean signal while removing the noise. The above results show that our method also outperforms the other methods in denoising waveforms.

(a)

(b)

(c)

(d)

(e)

(f)

3.2. Discussion

3.2.1. Discussion of the Results of the Model

In this section, we analyze the ECG denoising results for generative adversarial networks with different structures and Bi-LSTM networks. Figure 10 shows the denoised waveforms of one complete ECG cycle for the five networks. We can observe that GAN has the obvious cutting of R-waves in ECG waveforms, the denoised waveforms of DCGAN have low coincidence with the original clean signal, and LSTM-GAN is not effective enough for high-level noise reduction. Our method can not only remove some slight noise but also remove high-level noise. In addition, the denoised signal obtained by our method is smoother and has a higher coincidence with the original clean signal. We analyze the reasons for this result. Feature extraction is performed by the local convolutional kernel. Each convolution operation is performed on the receptive field, and each neuron only considers the information in the receptive field. A smaller receptive field corresponds to a smaller kernel size and stride, and more detailed features can be extracted. For the whole denoising process, deeper detailed information can be preserved. This is why many researchers choose smaller kernel sizes and stride sizes when designing convolutional neural networks, but convolutional kernels with smaller receptive fields are less capable of grasping global information. Therefore, the convolution-based signal denoising method is the reconstruction from local detailed information to global information. To consider global information, we start from two points: firstly, the adversarial strategy is employed. The adversarial approach begins with the overall data distribution to give the denoised signal access to the same distribution as the clean signal. Secondly, the global information is poorly preserved for the convolution of small size filters. The LSTM layer is added after each convolutional layer in our generative adversarial structure to associate the order dependence of the convolved features.

It can be seen from Figure 10 that the Bi-LSTM model can effectively remove the noise in the signal. To further verify the advantages and disadvantages of the two methods, we compare the denoising results of Bi-LSTM and LSTM-DCGAN for different kinds of noise (MA, BW, and EM) at various SNR. From the results in Table 3, Bi-LSTM outperforms our method in removing MA noise but performs worse in removing BW and EM noise. The SNR_imp of Bi-LSTM for the three types of noise removal ranges from 15.063 dB to 20.807 dB, and the difference in the removal of different noise reaches 5.744 dB, while the difference of our method is only 1.932 dB.

3.2.2. Robustness of Model Deployment

We choose the ECG signal in data set II to confirm the applicability of our proposed methods to other leads. MA+BW+EM mixed noise was added to the original V1 original signal at -1 dB, 3 dB, and 7 dB. Table 4 shows the results of the noise removal for the new leads. Figure 11 shows the waveform plot of the model’s denoising results for lead V1.

From the results in Table 4, the denoising results of V1 leads are slightly worse than the lead II leads at -1 dB and 7 dB and slightly better than the lead II at 3 dB. The differences in SNR_imp between the two leads at -1 dB, 3 dB, and 7 dB are 3.33%, 4.48%, and 9.67%, respectively. As a result, our method can be used to remove noise from ECG signals in other leads within an acceptable range.

The effectiveness of the model in denoising ECG signal differences in various individuals was tested. A 15-fold repeated experiment is designed with experimental data from the MIT-BIH Arrhythmia Database of 15 different patients; they are recorded as 100, 103, 106, 109, 115, 116, 123, 202, 205, 209, 220, 223, 230, 231, and 234. Of the 15 sets of records, 14 of them are used for model training, and the remaining 1 is used for model testing, thus repeating the experiment 15 times to observe the model’s ability to process “new data.” Since the ECG signals collected daily are often accompanied by various noise disturbances, the clean signals are polluted with a mixture of MA+EM+BW noise to form noisy ECG data.

The results in Table 5 show that the denoising effect of each lead varies due to individual differences, with the SNR_imp of the patient data of 115 records reaching 19.217 dB and the result of the patient data of 202 records being lower, only 11.900 dB. This is mainly caused by the data dependence of the model, which is unable to learn information about data that is not involved in training. To reduce the data dependency of the model, we use a small amount of new data for model boosting to obtain information about the new data. Model boosting is performed by adding a small number of samples to the existing model. Take the worst record 202 as an example, by adding 20% of the data (108 samples) into the model training and 80% (434 samples) as the test of the boosted model. The SNR_imp for lead 202 is increased from 11.900 dB to 15.237 dB.

4. Conclusions

We propose a GAN method using CNN combined with LSTM to draw attention to local and global information in ECG denoising studies. Convolution is used throughout the network structure to help the denoised signal preserve more detailed information. In order to make the global information highly preserved, we employ the generative adversarial strategy to guarantee that the data distributions of the original clean signal and the denoised signal are consistent. Additionally, we update the generator and discriminator of the GAN by adding an LSTM layer after each convolutional layer to provide time series information to the features. According to experimental results, our technique has a higher SNR_imp, a smaller RMSE, and a lower PRD than DCGAN and LSTM-GAN. The comparison of denoised waveforms also demonstrates that our method outperforms the other two models in terms of noise suppression and detailed preservation. However, our method requires a large amount of computation. Another important area for research in the future is lightweight denoising models, which are being used for portable wearable devices. In addition, there are many additional active and effective ECG-denoising models that need to be validated.

Data Availability

All the data utilized in our research can be accessed from https://physionet.org/content/mitdb/1.0.0/.

Conflicts of Interest

The authors reported no potential conflict of interest.

Acknowledgments

The authors would like to acknowledge each author for their suggestions and cooperation. This work was supported by the Natural Science Foundation for Young Scientists of Gansu Province, China (Grant No. 20JR5RA438); the Natural Science Foundation of Gansu Province, China (Grant No. 20JR5RA459); and the Foundation of Key Laboratory of Gansu Advanced Control for Industrial Processes (Grant No. 2022KX11).

References

Z. Shang, Y. Zhao, and Y. Lian, “Low power FIR filter design for wearable devices using frequency response masking technique,” in 2017 IEEE 12th International Conference on ASIC (ASICON),, pp. 516–519, Guiyang, China, 2017.
View at: Publisher Site | Google Scholar
S. S. Dhillon and S. Chakrabarti, “Power line interference removal from electrocardiogram using a simplified lattice based adaptive IIR notch filter,” in 2001 Conference Proceedings of the 23rd Annual International Conference of the IEEE Engineering in Medicine and Biology Society, vol. 4, pp. 3407–3412, Istanbul, Turkey, 2001.
View at: Publisher Site | Google Scholar
G. Lu, J.-S. Brittain, P. Holland et al., “Removing ECG noise from surface EMG signals using adaptive filtering,” Neuroscience Letters, vol. 462, no. 1, pp. 14–19, 2009.
View at: Publisher Site | Google Scholar
H. D. Hesar and M. Mohebbi, “An adaptive particle weighting strategy for ECG denoising using marginalized particle extended Kalman filter: an evaluation in arrhythmia contexts,” IEEE Journal of Biomedical and Health Informatics, vol. 21, no. 6, pp. 1581–1592, 2017.
View at: Publisher Site | Google Scholar
M. Izzetoglu, A. Devaraj, S. Bunce, and B. Onaral, “Motion artifact cancellation in NIR spectroscopy using Wiener filtering,” IEEE Transactions on Biomedical Engineering, vol. 52, no. 5, pp. 934–938, 2005.
View at: Publisher Site | Google Scholar
G. U. Reddy, M. Muralidhar, and S. Varadarajan, “ECG De-noising using improved thresholding based on wavelet transforms,” International Journal of Computer Science and Network Security, vol. 9, no. 9, pp. 221–225, 2009.
View at: Google Scholar
N. E. Huang, Z. Shen, S. R. Long et al., “The empirical mode decomposition and the Hilbert spectrum for nonlinear and non-stationary time series analysis,” Proceedings of the Royal Society of London Series A: Mathematical, Physical and Engineering Sciences, vol. 454, no. 1971, pp. 903–995, 1998.
View at: Publisher Site | Google Scholar
G. Singh, G. Kaur, and V. Kumar, “ECG denoising using adaptive selection of IMFs through EMD and EEMD,” in 2014 International Conference on Data Science & Engineering (ICDSE), pp. 228–231, Kochi, India, 2014.
View at: Publisher Site | Google Scholar
F. Xiao, D. Yang, X. Guo, and Y. Wang, “VMD-based denoising methods for surface electromyography signals,” Journal of Neural Engineering, vol. 16, no. 5, article 056017, 2019.
View at: Publisher Site | Google Scholar
S. Boda, M. Mahadevappa, and P. K. Dutta, “A hybrid method for removal of power line interference and baseline wander in ECG signals using EMD and EWT,” Biomedical Signal Processing and Control, vol. 67, article 102466, 2021.
View at: Publisher Site | Google Scholar
M. A. Kabir and C. Shahnaz, “Denoising of ECG signals based on noise reduction algorithms in EMD and wavelet domains,” Biomedical Signal Processing and Control, vol. 7, no. 5, pp. 481–489, 2012.
View at: Publisher Site | Google Scholar
S. A. Malik, S. A. Parah, and B. A. Malik, “A VMD-SWT based ECG denoising technique,” in 2021 Sixth International Conference on Image Information Processing (ICIIP), pp. 158–163, Shimla, India, 2021.
View at: Publisher Site | Google Scholar
Z.-G. Sun, Y. Lei, J. Wang, Q. Liu, and Q.-Q. Tan, “An ECG signal analysis and prediction method combined with VMD and neural network,” in 2017 7th IEEE International Conference on Electronics Information and Emergency Communication (ICEIEC), pp. 199–202, Macau, China, 2017.
View at: Publisher Site | Google Scholar
G. Wang, L. Yang, M. Liu et al., “ECG signal denoising based on deep factor analysis,” Biomedical Signal Processing and Control, vol. 57, article 101824, 2020.
View at: Publisher Site | Google Scholar
P. Xiong, H. Wang, M. Liu, and X. Liu, “Denoising autoencoder for electrocardiogram signal enhancement,” Journal of Medical Imaging and Health Informatics, vol. 5, no. 8, pp. 1804–1810, 2015.
View at: Google Scholar
H.-T. Chiang, Y.-Y. Hsieh, S.-W. Fu, K.-H. Hung, Y. Tsao, and S.-Y. Chien, “Noise reduction in ECG signals using fully convolutional denoising autoencoders,” IEEE Access, vol. 7, pp. 60806–60813, 2019.
View at: Publisher Site | Google Scholar
E. Dasan and I. Panneerselvam, “A novel dimensionality reduction approach for ECG signal via convolutional denoising autoencoder with LSTM,” Biomedical Signal Processing and Control, vol. 63, article 102225, 2021.
View at: Publisher Site | Google Scholar
J. Wang, R. Li, R. Li et al., “Adversarial de-noising of electrocardiogram,” Neurocomputing, vol. 349, pp. 212–224, 2019.
View at: Publisher Site | Google Scholar
B. Xu, R. Liu, M. Shu, X. Shang, Y. Wang, and J. P. Martínez, “An ECG denoising method based on the generative adversarial residual network,” Computational and Mathematical Methods in Medicine, vol. 2021, Article ID 5527904, 23 pages, 2021.
View at: Publisher Site | Google Scholar
F. Zhu, F. Ye, Y. Fu, Q. Liu, and B. Shen, “Electrocardiogram generation with a bidirectional LSTM-CNN generative adversarial network,” Scientific Reports, vol. 9, no. 1, pp. 1–11, 2019.
View at: Publisher Site | Google Scholar
I. Goodfellow, J. Pouget-Abadie, M. Mirza et al., “Generative adversarial nets,” Advances in Neural Information Processing Systems, vol. 27, 2014.
View at: Google Scholar
A. Radford, L. Metz, and S. Chintala, “Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks,” 2015, https://arxiv.org/abs/1511.06434.
View at: Google Scholar
S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural Computation, vol. 9, no. 8, pp. 1735–1780, 1997.
View at: Google Scholar
G. B. Moody and R. G. Mark, “The impact of the MIT-BIH arrhythmia database,” IEEE Engineering in Medicine and Biology Magazine, vol. 20, no. 3, pp. 45–50, 2001.
View at: Publisher Site | Google Scholar
G. B. Moody, W. Muldrow, and R. G. Mark, “A noise stress test for arrhythmia detectors,” Computers in Cardiology, vol. 113, no. 3, pp. 381–384, 1984.
View at: Google Scholar

Copyright

Copyright © 2023 Huidong Wang et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

PDF Download Citation

Download other formats

Order printed copies