Intelligent approach for artifacts removal from EEG signal using heuristic-based convolutional neural network

Highlights

• Exploits SM-EFO to automatically remove noise from the contaminated EEG signal.

• Promotes a SM-EFO-based optimization for the efficient removal of artifacts.

• The performance of the artifacts removal model has shown better performance.

Abstract

In general, the electrical activity of the brain is recorded using Electroencephalography (EEG), which is contaminated with some signal artifacts. By using automatic removal of artifacts from EEG signals, different Brain-Computer Interface (BCI) and clinical diagnostics applications are in practice. However, they are not efficient to remove the artifact from the EEG signal. Thus, we plan for the intelligent model for artifacts removal of EEG signal. The two main phases of the proposed model are training and testing. The deep learning model in the training phase is used as the filter to automatically remove noise from the contaminated EEG signal. The proposed model adopts improved One-Dimensional Convolution Neural Networks (1D-CNN) for artifacts removal from EEG signals. Here, a new hybrid algorithm named Spider Monkey-based Electric Fish Optimization (SM-EFO) is proposed by integrating the Spider Monkey Optimization (SMO) and Electric Fish Optimization (EFO) algorithm. The model parameters of the One-Dimensional Convolutional Neural Networks (1D-CNN) are tuned by using SM-EFO. The experimentation is performed on the standard benchmark dataset, and the experimental results establish that the proposed model can achieve significant improvement and get cleaner waveforms in terms of several performance measures when compared to the conventional models.

Introduction

EEG records the electrical activity of the cerebral cortex by electrodes, which are generally installed on the scalp. This approach is commonly employed for the medical diagnosis of sleep disorders and epilepsy [1]. Nowadays, EEG attracts more medical practitioners in the field of BCI applications [2]. Though, the EEG signals are generally distorted due to the environmental and biological artifacts for practical settings [3]. Undesired signals [4] are considered artifacts that include non-cerebral origins recorded by EEG electrodes [5]. The artifacts are categorized into environmental [6] artifacts and biological artifacts [7]. Environmental artifacts are generally originated from the external part of the human body owing to interference or electrode movement from external devices like electric motor or power main [8] whereas the signals from non-cerebral sources in the human body like muscle or ocular and cardiac activity are considered as biological artifacts [9], [10]. However, both environmental and biological artifacts degrade EEG signals, and thus BCI applications or medical diagnosis [11] is obstructed through the distortion of the considered power spectrum [12].

The consistent interpretation of the EEG recordings is influenced by the limitations of muscle artifacts, and hence a new emerging case is developed by adopting the wearable few-channel EEG [13]. Though, the low robustness and high computational load of the conventional approaches restrict their performance and broader applications in artifacts removal. When the EEG data is collected from the recording systems, the most important scenario is signal artifacts [14], [15], which may also contaminate the quality of EEG data. Moreover, complete knowledge of the artifacts types is required for efficiently removing the noise or artifacts. The unwanted signals or artifacts are mostly derived from physiological artifacts, experimental error, and environment noise [16]. In addition, the external factors such as experiment error and environment artifacts are categorized as extrinsic artifacts, where the intrinsic artifacts are taken from physiological activities like muscle activity, heartbeat, eye blink, etc from the body [17], [18]. The most important physiological artifacts such as extrinsic artifacts, cardiac artifacts, muscle artifacts, and ocular artifacts affect the EEG data [19].

Recent researchers have used different Time-Frequency Representation (TFR)-based signal decomposition approaches for removing the EEG artifacts [20]. Some of the TFR techniques like Wavelet Transform (WT), Short Time Fourier Transform (STFT), etc are proposed for extracting the fine-scale fluctuations in EEG signals [21]. The pure EEG activity is attained when the inverse WT of threshold wavelet coefficients is applied, which is completely based on EEG de-noising approaches. Various methods are developed for the removal of EMG artifacts, where the popular methods like signal decomposition approaches and Blind Source Separation (BSS) techniques such as ICA and Canonical Correlation Analysis (CCA) are chosen. Here, both BSS methods are efficiently used for removing the artifacts from the experimentation results, which are validated regarding computation time and accuracy [22]. Ensemble Empirical Mode Decomposition (EEMD) [18] attains more benefits concerned with their better capability of processing the non-stationary signals while considering the signal decomposition methods that were also avoided to select the wavelet basic in WT based on the data-driven principle [22]. The major issue of the EEG artifacts removal is the selection of the threshold level because it will never maintain the artifacts signals as the original ones, and it does not eliminate the coefficients of the original EEG signal [23]. Therefore, there is the necessity of designing a novel artifacts removal model.

The major contribution of the developed model is given here.

• To propose a new artifacts removal model from EEG signal by proposing a heuristic-based 1D-CNN using a hybrid optimization algorithm.

• To propose a new hybrid SM-EFO algorithm for efficient artifacts removal model by developing an improved 1D-CNN through the optimization of the number of hidden neurons and number of epochs to maximize the multi-objective function.

• To propose a new improved 1D-CNN using hybrid SM-EFO algorithm for attaining the clean signal from the 400 × 1 window of the noisy signal by minimizing the multi-objective function concerned with PSNR and RMSE.

• To analyze the performance of the developed artifacts removal model of EEG signal with different ECG, EMG, and EOG artifacts in terms of different measures while comparing with some meta-heuristic-based algorithms.

The remaining sections of the developed model are given here. Section II explains the related works. Section III proposes an enhanced deep learning-based artifacts removal from EEG signal. Section IV describes the improved 1D-CNN for artifacts removal from EEG signal. Section V implements the SM-EFO algorithm for improving the 1D-CNN architecture. Section VI analyzes the results. Section VII concludes this paper.