Two Class Motor Imagery EEG Signal Classification for BCI Using LDA and SVM

The strategy and materials used in the proposed work are illustrated in Figure 1. The data were collected from the BNCI Horizon 2020 website, specifically from BCI Competition IV dataset 2a [16]. Each subject's data includes 22 EEG channels and 3 EOG channels, with the left mastoid serving as the reference. Only the EEG channels were used for processing in the proposed work. The collected MI EEG information is shipped off channel determination block where the prevailing channels were recognized from 22 MI EEG channels. The dominant channels were identified using ECTA. In ECTA method the energy of each channel for the 3-seconds window is estimated and which is compared with the 60% of the maximum energy of that particular channel. If the energy value for the 3-seconds window is greater than the threshold value then the count value for that particular channel will be incremented by one. This procedure is repeated for the all the 22 channels and as a result 9-channels were identified as the dominate. These dominant channels are handled utilizing wavelet decomposition method and decomposed into four levels in the signal processing stage as mentioned in the Figure 1. From the extracted rhythms Mu and Beta band is separated. From these two bands Energy and Entropy feature values were extracted from every dominant channel. From the extracted features, 80% of the data was used for training, while the remaining 20% was used for testing classifier models. The results obtained were analyzed and compared with existing methods.

1.png

Figure 1. Block diagram of proposed work

2.1 Data collection

The data used in this study was collected from the BNCI Horizon 2020 website, specifically from the BCI Competition IV dataset 2a [16]. This dataset comprises MI EEG signals from nine healthy subjects, each participating in two sessions: a training session and a test session. The recorded tasks include the imagination of movements of the left hand (Class 1), right hand (Class 2), both feet (Class 3), and tongue (Class 4). Each subject's data includes 22 EEG channels and 3 EOG channels, with the left mastoid serving as the reference. Each subject participated in two recording sessions, each consisting of six runs separated by short breaks. Each run included 48 trials (12 trials per each of the four classes), resulting in a total of 288 trials per session.

During the experiment, subjects were seated comfortably in an armchair facing a computer screen. At the beginning of each trial (t = 0 s), a fixation cross appeared on the black screen, accompanied by a brief acoustic warning tone. After two seconds (t = 2 s), a visual cue in the form of an arrow pointing left, right, down, or up appeared on the screen for 1.25 seconds. Subjects were instructed to perform the motor imagery task corresponding to the cue until the fixation cross disappeared.

2.2 Dominant channel selection

One of the major concerns with BCI applications is that when they are implemented in real-time, the number of electrodes used to acquire MI EEG signals must be kept as low as possible so that the subject feels comfortable and can perform well in the activities. Therefore, this in turn helps in getting better MI EEG signals so that these signals can be classified easily and converted into action commands for different applications. So, in the proposed work, we implemented a methodology named ECTA that aids in reducing the number of MI EEG channels. The implementation of this proposed method uses the following steps:

Step 1: Find the maximum energy of each dominant channel for the total duration of the MI EEG signal.

Step2: For every 3-sec window, energy values for each channel were extracted.

Step 3: If the extracted energy value is greater than 60% of the maximum energy of that channel, then increase the energy count value by one (The initial energy count value is zero).

The mathematical condition for the above step is represented as follows:

If $\sum_{n=1}^{22} E_{\text {channel_n }}>\left(0.6 * E_{\max \_c h a n n e l_{-} n}\right)$ - then increase the energy count value for the channel number ‘n’. Where, $E_{\text {channel_n }}$ – Energy of the channel number ranging from 1 to 22 for three seconds window. $E_{\max \_c h a n n e l_{-} n}$ – Maximum Energy of the channel number ranging from 1 to 22 for total duration of the MI EEG signal.

Table 1 shows the computed energy count for each channel of subject-1, and it can be seen that the energy count ranges from 0 to 46. Based on the energy count value, nine dominant channels were identified. The channels with energy count ranging from 30 to 46 were identified as dominant channels. The remaining eight subjects were analyzed in a similar manner. Based on the ECT approach from the Frontal, Central, and Parietal lobes, nine channels (Fc1, Fcz, Fc2, C3, Cz, C4, P1, Pz, and P2) were identified as the dominating channels for each subject. In the Figure 2 the number (3, 4, 5, 8, 10, 12, 19, 20, and 21) in bold circle represents the dominant channel and other numbers (1, 2, 6, 7, 9, 11, 13, 14, 15, 16, 17, 18, and 22) are rejected channels. Although several of the Centro parietal channels (CP3, CP1, CPz, and CP2) are dominant, they are not employed for further study because channels from the Central and Parietal lobes have already been picked for the feature extraction. In addition, channel 22 is discarded because the signal level of the Occipital region can be altered by visual impacts. As described in the Data Collection section, the MI EEG recordings were conducted with subjects seated comfortably in an armchair in front of a computer screen. The motor imagery task was prompted by visual cues in the form of arrows pointing left, right, down, or up, which appeared on the screen and remained for 1.25 seconds [16]. Therefore, as this visual effect can affect signal intensity in channels from the occipital lobes, channels from this lobe are not examined for feature extraction. The proposed method of selecting dominating channels reduces computational complexity while simultaneously could provide significantly more comfort to the subject during real-time recording and implantation of proposed work.

2.png

Figure 2. 10-20 Electrode system

Table 1. Energy count of each channel

2.3 Discrete wavelet transform

MI EEG signals are non-stationary signals in the traditional sense. Short-time Fourier transforms (STFTs) were used to analyze EEG data as a time-frequency analysis method, although it was found that the window selection is crucial for the short-time Fourier transform. The wavelet transform, which is a multi-resolution analysis approach, provides a solution to this problem and could provide a more precise temporal localization. For non-stationary signals, the Wavelet transform is a new two-dimensional time-scale processing approach [17]. Its primary benefit is that it provides simultaneous information on the frequency and temporal location of signal characteristics in terms of signal representation at several resolutions corresponding to various time scales. The Discrete Wavelet Transform (DWT) produces two coefficients, Di and Ai, which are the down sampled outputs of the high-pass and low-pass filters at each decomposition level [18]. A four-level decomposition employing a Daubechies wavelet of order 4 (db4) is applied to the nine dominating MI-EEG signals. Because db4 wavelet stands as a widely employed choice for EEG signal decomposition due to its advantageous properties such as compact support, orthogonality, and vanishing moments, facilitating effective analysis and feature extraction of both transient and oscillatory components in EEG signals. Higher-order mother wavelets, possessing more vanishing moments, offer increased localization in both time and frequency domains, potentially improving resolution in signal decomposition and enabling finer detail capture in MI-EEG signal dynamics. Two rhythms, the Mu band and the Beta band, were extracted from MI-EEG data using the wavelet decomposition approach.

In the flow diagram of Figure 3, the approximated and detailed coefficients of the MI-EEG signals utilizing four-level wavelet decomposition are shown. The Energy and Entropy features were derived from the wavelet coefficients (Cad3 and Cad4), which had a frequency range of 15.6Hz to 31.2Hz for Cad3 and 7.81Hz to 15.6Hz for Cad4, as shown in the Figure 3.

3.png

Figure 3. Frequency representation of approximated and detailed coefficients of MI-EEG signals

2.4 Feature extraction

In BCI literature, the most commonly used features in MI-based BCI applications for the classification of multiple MI tasks are power spectral density and energy of Mu and delta band. In the proposed work for all the 9-subjects two different features namely, energy and entropy from Mu and Beta bands were extracted for 288 trials per session from a 3-seconds window. The features were extracted only from 9-dominant channels (Fc1, Fcz, Fc2, C3, Cz, C4, P1, Pz, and P2) using the following mathematical expressions:

Energy $=\sum_{i=1}^n S_i^2$               (1)

where, S is the MI-EEG signal.

Entropy $=-\sum_{k=1}^n P_k * \log _2\left(p_k\right)$        (2)

where, P_k is the probability of occurrence of k^th EEG sample.

The extracted features from all the nine subjects were used for testing and training both LDA and SVM classifiers. Eighty percent of the features were used to train the classifiers, while twenty percent of the data was utilized to test them. The resulting results were compared to existing approaches. The performance of each subject with both the classifiers is explained in the next section. In the following section, we will look at how each subject performed with both classifiers.

2.5 Classification

In BCI literature, LDA and SVM are frequently used for categorizing MI EEG tasks, often achieving better classification accuracy compared to other models. SVM, in particular, is regarded as a highly promising tool for classifying single EEG trials [13, 14]. Ma et al. [14] highlighted that SVM effectively addresses issues related to small sample sizes and high-dimensional data, resulting in improved classification accuracy.

LDA and SVM are also used in the proposed method to classify two classes of MI EEG data. The extracted features from the data set are loaded into excel sheet. The excel sheet file with the energy and entropy features were fed as input to the different classifiers using Classification Learner App in MATLAB software. This App in MATLAB offers the advantage of intuitive and interactive model building for classification tasks, simplifying the process for users without extensive programming or machine learning expertise.

LDA is a popular supervised learning method that is widely used for dimensionality reduction as well as classification tasks, particularly in the context of two-class classification situations (Class1, Class4). Through statistical modeling of feature distributions and subsequent computation of an optimal decision boundary that maximized between-class distance while minimizing within-class variance, LDA worked by identifying the best linear combination of proposed features that effectively separated the classes (Class1 and Class 4). This illustrated a linear border in the feature space and is the quintessential LDA. Specifically, LDA is a strong technique that works well for binary classification problems where classes can be clearly defined by a linear border. When faced with new data, LDA allocated class membership based on the highest posterior probability relative to this decision boundary.

In the proposed work, an SVM model with a linear kernel function is utilized. Using a linear kernel in SVM is a powerful method for binary classification tasks. SVM maximizes the margin, which is the distance between the hyperplane and the closest data points, or support vectors, by identifying the optimal hyperplane in the feature space. By calculating dot products between feature vectors, the linear kernel function in SVM effectively transforms the data into a higher-dimensional space where class separation is feasible. The objective is to identify a hyperplane that minimizes the risk of overfitting while effectively classifying data and generalizing well to unseen data. New data points are classified based on their position relative to the hyperplane. In summary, SVM with a linear kernel provided a reliable and effective solution by determining the best decision boundary.