Electroencephalogram (EEG) is susceptible to various nonneural physiological artifacts. Automatic artifact removal from EEG data remains a key challenge for extracting relevant information from brain activities. To adapt to variable subjects and EEG acquisition environments, this paper presents an automatic online artifact removal method based on a priori artifact information. The combination of discrete wavelet transform and independent component analysis (ICA), wavelet-ICA, was utilized to separate artifact components. The artifact components were then automatically identified using a priori artifact information, which was acquired in advance. Subsequently, signal reconstruction without artifact components was performed to obtain artifact-free signals. The results showed that, using this automatic online artifact removal method, there were statistical significant improvements of the classification accuracies in both two experiments, namely, motor imagery and emotion recognition.
As a biological signal that reflects potential changes in complex brain activities, electroencephalogram (EEG) plays an important role in human brain research, disease diagnosis, brain-computer interfaces (BCI), and so on. However, the electrical signals of brain activities are weak, so real EEG is susceptible to various nonneural physiological artifacts. The most severe artifacts include eye movement (electrooculography, EOG) and muscle movement (electromyography, EMG) artifacts [
Artifact avoidance and artifact rejection were used to handle artifacts in early studies. These approaches might not acquire sufficient valid data from actual experiments, in which eye blinking, swallowing, or other nonneural physiological activities are inevitable [
Blind signal separation (BSS) techniques are the most promising approach for separating the recordings into components that “build” them. They regard EOG, EMG, and other artifacts as the signals produced by independent sources. BSS techniques need to identify components that are attributed to artifacts and perform signal reconstruction without them [
Automatic artifact removal from EEG is preferred in practice. It is suitable for only EOG artifact removal with a reference channel [
This paper proposes a novel automatic artifact removal method for variable subjects and EEG acquisition environments. Without reference channels and massive offline training samples, a small amount of time is used to acquire individual artifact samples as online a priori artifact information in advance. Automatic identification and removal of artifact components are realized using correlation analysis and wavelet-ICA (WICA). At last, the method is applied to two classification experiments, namely, motor imagery and emotion recognition. The experimental results showed that there were statistical significant improvements of the classification accuracies by applying this automatic online artifact removal method.
The following subsections describe how the proposed automatic artifact removal approach was established. We also applied the approach to two classification experiments, namely, motor imagery and emotion recognition.
We first describe how to obtain a priori artifact information online, which is necessary for the following automatic artifactual component identification. During the actual EEG acquisition, artifacts are often generated by the movements of subjects, intentionally or unintentionally, such as eye blinking, eye rolling, teeth clenching, and swallowing. If the subject does only one action for a time period, the corresponding recoding data can be clearly marked by a corresponding artifact label, which can be utilized for the following automatic artifact classification. Thus, an artifact acquisition session was performed to extract a priori artifact information before the formal EEG data acquisition. Figure
One trial of the artifact acquisition session. One trial consists of one 1 s blank period, one 1 s ready period, one 2 s action period, and one 2 s rest period. When a visual cue is presented in the action period, the subjects are required to do the corresponding action only. The EEG epoch represents the data used for the following automatic artifact removal.
As a combination of DWT and ICA, WICA is proposed based on the joint use of multiresolution and multidimensional analyses. WICA was first proposed for the processing of EMG signals [
This paper presents a novel WICA technique optimized for automatic EOG and EMG artifact removal using individual a priori artifact information acquired online in advance. This section explains how the proposed method allows automatic EOG and EMG artifact removal.
The algorithm for EOG and EMG artifact removal in EEG (as shown in Figure
Block diagram of WICA for automatic EEG artifact removal. Raw data to be removed are appended to the artifact samples first. The next stage is wavelet decomposition via channel by channel, in which data are projected into
(1) Raw data to be processed were appended to the artifact samples first. Mallat’s pyramid decomposition algorithm was applied to
(2) All coefficient vectors
(3) The a priori artifact information and correlation analysis were applied to recognize EOG and EMG wavelet-independent components (WICs). First, Mallat’s pyramid construction algorithm was applied to each WIC
(4) Entity matrix
(5) Mallat’s pyramid construction algorithm was applied to each channel of wavelet coefficients
In this section, we applied our automatic artifact removal approach to a motor imagery classification experiment. We investigated the effects of our approach on the classification performance of motor imagery based on PSD.
Fourteen healthy BCI novices performed first motor imagery with the left hand, right hand, and neither in a calibration measurement without feedback. Every 10 s, one of three different visual cues (arrows pointing left, right, or both) indicated to the subject which type of motor imagery to perform (Figure
EEG recording parameters.
EEG recording parameters | |
|
|
Amplifier | 16-channel g.USBamp system (gtec, Graz, Austria) |
Sampling frequency | 512 Hz |
High-pass filter | 0.1 Hz |
Low-pass filter | 60 Hz |
Notch filter | 50 Hz |
Electrode placements | 16-channel subset of 10–20 systems (see Figure |
Ground | Forehead |
Reference | Right earlobe |
Electrode material | Ag/AgCl |
Recording software | g.Recorder |
One trial of the motor imagery experiment. The EEG epoch represents the data used for analysis and classification.
After artifact removal, data from three electrodes (C3, CZ, and C4) in the motion imagery period were selected to extract the power spectrum feature for the input of SVM classification. First, the 4 s long epoch was equally divided into four segments. Second, each segment of the EEG data was processed with the Hanning window. Third, windowed segments were extended by zero padding for fast Fourier transform. Finally, EEG power spectra were extracted in 45 bands from 1 Hz to 45 Hz, and each band was 1 Hz long. Thus, the total number of feature dimensions was 540.
Correlation-based feature selector is a type of supervised dimensionality reduction method [
To evaluate the artifact removal performance for EEG data from higher-order cognitive processes, data from thirteen healthy subjects were used to test the proposed automatic artifact removal method in another classification experiment, namely, an emotion recognition study.
Thirteen college participants aged 20 to 24 years with normal or corrected-to-normal vision participated in this study. All participants had no neurological or psychological medical history. Before experiments, we obtained informed consent from each participant. The pictures used for emotion induction were obtained from the Chinese Affective Picture System [
The emotion induction experiment (150 trials) is illustrated in Figure
One trial of the emotion recognition experiment. The EEG epoch represents the data used for analysis and emotion classification.
Power spectrum features are widely used for emotion recognition because they can be analyzed to characterize the perturbations in the oscillatory dynamics of ongoing EEG [
In the following subsections, we investigated the validity of our method via correlation analysis for identifying artifact components. We also comparatively analyzed the signal waveform and classification performances of two validation experiments before and after artifact removal.
We plotted the corresponding time-domain components of the WICs to visually inspect and identify the artifacts and compared the results of automatic identification by correlation analysis. Figure
Corresponding time-domain components of the WICs.
Correlation scores
From another perspective, the different distributions of correlation scores between EOG and EMG may indicate the different characteristic power spectrum between them. EOG artifact components showed significantly higher energy in low power spectrum (1–10 Hz) (Figures
We implemented our method in Matlab 2012b. The average computation cost of the automatic artifact removal method for one single trial was about 5 s. In it, The DWT and ICA accounted for a higher proportion (approximately 4.95 s). Given that the shortest single trial of our validation experiments was 10 s, our method met the requirements of real-time analysis.
Figure
Raw EEG signals with strong EOG and EMG artifacts.
Figure
Artifact-removed EEG signals (signals correspond to those depicted in Figure
All the classification tests in this study were carried out using fivefold cross validation with RBF-kernel SVM. We calculated the offline classification accuracies with different numbers of features selected by different correlation score thresholds. For all subjects, we compared the highest accuracies of the classification between raw data and artifact-removed data. Both the results of binary-category (left and right) classification and three-category (left, right, and neither) classification were utilized to test our method. The mean highest accuracy of binary-category classification across fourteen subjects is shown in Figure
Classification accuracies of raw data and artifact-removed data for binary-category (left and right) and three-category (left, right, and neither) classification. For each subject, an appropriate number of features were selected for the highest accuracy. The mean accuracy was computed across all the subjects. Error bars show the standard deviation of the mean accuracies across all subjects.
In Validation 2, we performed a binary-category (VHV+HV and LV+VLV) classification and five-category (VHV, HV, neutral, LV, and VLV) classification to verify our method. In contrast to Validation 1, features were selected from all the 16 channels, and the maximum number of features was 2880. We also compared the highest accuracies of the classification performances between raw data and artifact-removed data among all subjects (Figure
Classification accuracies of raw data and artifact-removed data for binary-category (VHV+HV and LV+VLV) and five-category (VHV, HV, neutral, LV, and VLV) classification. For each subject, an appropriate number of features were selected for the highest accuracy. The mean accuracy was computed among all the subjects. Error bars show the standard deviation of the mean accuracies across all subjects.
The main idea of our method was to acquire the online a priori artifact information and put them in the WICA with the raw data including artifact to be removed. The artifact components were recognized and removed by sorting the correlation of the marked a priori artifact information and WICs. The a priori artifact information obtained online can effectively reflect the nonneural physiological artifacts during the EEG experiments. And the types of artifacts to be removed were determined by the types of the a priori artifact information. We used eye blinking, eye rolling, and teeth clenching to generate a priori EOG and EMG artifact information, respectively. The EOG artifact produced by eye blinking and eye rolling was mainly contained in two ICs (Figures
There is one thing that needs to be stressed which is that the performance of the chosen ICA method directly determines whether the artifact components can be separated. In this study, we chose WICA by analyzing different methods during preexperiments. Compared with general ICA methods, WICA improves the performance of ICA, since it projects data into a new space where the redundancy is higher and the features of artifacts are fully utilized. The statistical results also demonstrated that it was effective for motor imagery and emotion recognition. However, it cannot be ruled out that other ICA methods may work well in different conditions. Since we only focused on the automatic online artifact removal method in this study, we did not do much research in feature extraction and classification methods, which might affect the classification accuracies more or less.
In this study, a priori artifact information acquired online was introduced into WICA to realize automatic artifact removal for variable subjects and EEG acquisition environments. The proposed method was applied to two experiments, namely, motor imagery and emotion recognition. The statistical results showed that our method significantly improved the classification accuracies for motor imagery and emotion recognition. In addition, our method required no reference channels, massive training samples, and visual inspections, so it was entirely automatic. Therefore, the proposed method may provide an alternative approach for automatic artifact removal, particularly for novice researchers in other fields.
The authors declare that there is no conflict of interests regarding the publication of this paper.