A novel end-to-end 1D-ResCNN model to remove artifact from EEG signals

Abstract

Electroencephalography (EEG) signals are an important tool in the field of clinical medicine, brain research and the study of neurological diseases. EEG is very susceptible to a variety of physiological signals, which brings great difficulties to the research and analysis of EEG signals. Therefore, removing noise from EEG signals is a key prerequisite for analyzing EEG signals. In this paper, a one-dimensional residual Convolutional Neural Networks (1D-ResCNN) model for raw waveform-based EEG denoising is proposed to solve the above problem. An end-to-end (i.e. waveform in and waveform out) manner is used to map a noisy EEG signal to a clean EEG signal. In the training stage, an objective function is often adopted to optimize the model parameters and in the test stage, the trained 1D-ResCNN model is used as a filter to automatically remove noise from the contaminated EEG signal. The proposed model is evaluated on the EEG signal from the CHB-MIT Scalp EEG Database, and the added noise signals are obtained from the database. We compared the proposed model with the independent of the composite analysis (ICA), the fast independent composite analysis (FICA),Recursive least squares(RLS) filter,Wavelet Transform (WT) and Deep neural network(DNN) models. Experimental Results show that the proposed model can yield cleaner waveforms and achieve significant improvement in SNR and RMSE.Meanwhile, the proposed model can also preserve the nonlinear characteristics of EEG signals.

Introduction

Electroencephalography (EEG) is the electrical response of brain cells in the cerebral cortex. It is typically collected by an electrode collection system placed on the head of the brain [1]. Through the analysis of EEG, we can obtain rich physiological, psychological and pathological information. The EEG signal not only shows brain function but also shows the state of all body systems. In addition, the EEG signal plays an important role in the detection and treatment of brain diseases such as epilepsy and brain tumors and can be used to diagnose brain death [2], [3]. However, EEG is a highly random nonlinear non-stationary signal, which contains very complex components, and the signal amplitude is microvolts, the intensity is very weak, and it is very susceptible to other physiological signals of the human body, such as ElectroOculogram(EOG), Electrocardiogram (ECG), muscle artifacts(EMG), or interference from non-physiological signals such as spatial electromagnetic noise. These artifacts exist in almost the entire EEG acquisition process and often mask the waveform characteristics of EEG, which makes the reading of EEG signals more difficult and brings great difficulties to the subsequent research and application of EEG signal [4], [5]. Therefore, it is of great theoretical and practical significance to develop relevant methods to remove artifacts from EEG signals and remain useful information.

The most common method of denoising EEG is the artifact elimination method, which is to identify and remove artifacts in the brain signal and to completely preserve the neurological features and phenomena of the original signal [6], [7]. Artifacts reduction algorithms are mainly done in two ways: The first is through regression and filtering methods; the second is by separating or decomposing EEG data and noise data into other domains [8].

The regression model uses a function to fit the data to smooth the data. Using a regression analysis method, observe the multimodal linear model or nonlinear model between each EEG signal channel and between the EEG signal channel and the reference signal EOG, EMG or ECG channel, and find the mathematical equation suiTable for the data to eliminate noise [9], [10], [11]. However, this method only works for reference channels that are available.

In digital signal processing, the filter is an important unit. Adaptive filters can be divided into linear and nonlinear adaptive filters. Nonlinear adaptive filters further include Volterra filters and neural network based adaptive filters [12], [13], [14], [15], [16], [17]. Nonlinear adaptive filters have stronger signal processing capabilities and complex calculation [17], [18]. Linear adaptive filters are too sensitive and unsTable to adjust the parameters. Most importantly, artifacts overlap most of the clean EEG signals [19], [20]. Therefore, filters may eliminate useful EEG signals during artifact deletion.

Theempirical mode decomposition(EMD) proposed by Huang [21] decomposes the input signals into multiple empirical modes according to the inherent mode (IMFs) function, which is beneficial to the analysis of multi-component signals.EMD is an empirical and data-driven method for dealing with non-stationary, nonlinear, stochastic processes, so it is well suited for EEG signal analysis and processing.However,EMD is computationally complex and may not be suiTable for online applications [22], [23].

Blind source separation (BSS) is one of the most popular artifact removal methods [24], [25], [26], [27]. Independent Component Analysis (ICA) is a multi-dimensional signal processing method developed from BBS, which can separate the ideal signal and noise included in the EEG signal as independent components to achieve EEG signal denoising [28], [29]. Many BSS algorithms require human intervention to identify artifact components. This makes it subjective and time consuming [28], [29].

The wavelet transform(WT) maps the signal to the wavelet domain. According to the wavelet coefficients of signal and noise, they have different properties and mechanisms at different scales, eliminating the wavelet coefficients generated by noise and maximally retaining the coefficients of real signals [30], [31], [32].

EEG signal is a complex chaotic signal with nonlinear characteristics, and the preservation of nonlinear features is of great significance for EEG analysis and classification. Therefore, an EEG noise reduction method is needed to make the denoised EEG signal still maintains non-linear characteristics. On the other hand, with the advancement of technology, the collection of EEG data has become more convenient, providing a strong support for the implementation of deep learning. In recent years, deep learning techniques have been able to learn high-level and hierarchical representations directly in massive raw data, thus achieving a series of breakthroughs in signal processing. Specifically, Xu et al. [33] applied Deep Neural Networks(DNN) to speech enhancement, and proposed a DNN-based minimum mean square error regression fitting speech enhancement algorithm based on logarithmic power spectrum of the complex relationship between noisy speech and clean speech. Xu et al. [33] and Rodrigues and Couto [34] introduces an Restricted Boltzmann Machine(RBM)-based ECG denoising method, and [35] proposes an improved denoising automatic encoder (DAE) improved by wavelet transform (WT) for ECG denoising.Yang [36] introduced a Deep Learning Networks(DLN) EEG denoising method, which subtly utilizes the structural features of deep learning and powerful learning capabilities to improve EEG denoising due to EOG artifact.

Convolutional Neural Networks(CNN) is a subset of deep learning, which has attracted a lot of attention in recent years [37], [38]. It has been used in other fields for raw continuous signals, and it works well, starting with image applications, followed by many other fields, such as natural language processing, speech processing [39]. In contrast, CNN networks have not yet been used in remove artifacts from EEG signals. However, CNN networks have recently found applications in studies focusing on other areas of EEG time series analysis. In works [40], [41], [42] CNN is implemented on EEG signals, and the effectiveness of CNN algorithm in signal analysis is studied. In general, CNN has several natural advantages.First, CNN is superior to traditional methods not only in accuracy but also in speed.Second,CNN can automatically extract and learn the best features from the original signal to achieve adaptive design.The most important point is that CNN is good at mining the spatio-temporal structure in the input EEG signal, with weight sharing and deformation robustness.Although they have inherent advantages in EEG analysis, the CNN model has not received sufficient attention in EEG noise reduction.

EEG signals are usually long one-dimensional complex signals.Due to the time-varying and complexity of EEG signals, more complex nonlinear features should be extracted for EEG artifact removal.Therefore,in order to overcome the shortcomings of the above traditional methods,and considering the nonlinear characteristics of EEG time-varying signals and the advantages of CNN feature extraction, this paper proposes a new one-dimensional residual convolutional neural network model (1D-ResCNN) based on multi-scale kernel to remove noise from the EEG signal.The proposed model can automatically learn the nonlinear and discriminative deep features of the noisy EEG data and true EEG data.Then, these features are used to distinguish them and automatically reconstruct to obtain a clean EEG signal.

The main contributions of this paper are as follows:

• In the absence of sufficient prior knowledge, a new 1D-ResCNN based brain signal denoising model is proposed, which is the first application of CNN in EEG denoising;

• The proposed model operates directly on the raw EEG signal without pre-processing or manual feature extraction.Mo-reover, the EEG nonlinear characteristics and the details of the waveform are also kept;

• Combining the residual blocks of different scales, the model obtains more abundant features and increases the nonlinear expression ability of the convolutional neural network.

The structure of this paper is as follows. Section 2 describes the structure of the proposed 1D-ResCNN; Section 3 gives the experimental settings and datasets; Section 4 gives the experimental results; Finally, Section 5 shows the conclusion.