EEG signals are the overall reflection of the physiological activities of brain nerve cells in the cerebral cortex and scalp. By classifying and processing EEG signals, it is possible to identify states that do not require conscious activity. This article mainly processes the raw data and uses the multi-layer perceptron (MLP) neural network to determine whether the subject's eyes are open or closed and compares the results of the convolutional neural network (CNN) network.
Brain-Computer Interface (BCI) is a communication control system established
between the brain and external devices (computers or other electronic devices)
through signals generated during brain activity [
This study created a convolution neural network that can recognize and automatic extract the features of EEG signals and compare the accuracy of traditional methods of feature extraction and classification using data from the same public database. We used PhysioNet EEG data for this project, which are composed of over 1500 one- and two-minute EEG recordings, received from 109 subjects. The goal of our work is to explore Fast Fourier Transform (FFT) signal analysis techniques for distinction between two states, eyes open (EO) and eyes closed (EC), through the detection of EEG activity obtained from eight scalp channels.
PhysioNet EEG data set are composed of over 1500 one- and two-minute EEG
recordings, received from 109 subjects [
The 64-channel EEG were recorded (each sampled at 160 samples per second) as per the international 10-10 system (excluding electrodes Nz, F9, F10, FT9, FT10, A1, A2, TP9, TP10, P9, and P10), as it is shown Figure 1.
FT8 40 T8 42
T10 44 T9 43
T417 TP7 45 F3T97 FC15 FC23 FC31 FC4Z
AF7 25
A trained neural network will be planning to use to implement real-time classification in the future by our own device, which is an Ultracortex Mark IV biosensing headset from OpenBCI. In this case, only these eight channels were taken into account that are C3, C4, Fp1, Fp2, P7, P8, O1, and O2.
Since the data from eight channels original scalp are around two minutes long with 160 samples per second sampling frequency. In order to increase the number of samples, we expanded the data cut every 4 seconds as a segment (640 points).
EEGLAB is an interactive MATLAB toolbox for processing continuous and
eventrelated EEG, MEG, and other electrophysiological signals [
EEG signals can be analyzed in the frequency domain. We name 8-14 Hz for
alpha, 14-30 Hz for beta, 30-80 Hz for gamma, 1-4 Hz for delta, 4-8 Hz for theta
band, however frequency ranges are lightly dissimilar in various articles [
Figure 4 and Figure 5 respectively showed the activity spectrum at O2 position
of EC and EO state. We can see a distinct peak at 8-14hz on Figure 4 compared
to Figure 5. Alpha waves appear when people are awake, quiet and with their eyes
closed. As soon as subjects open their eyes, think, or receive other stimuli, alpha
waves disappear and turn into fast waves. It reappears when the person becomes
quiet again and closes his eyes. This phenomenon is called “alpha blocking” [
At the same time, alpha wave activities are unstable according to the Figure 4 and 5. When analyzing the same EEG signal, the amplitude of the beta waves are much less than that of the alpha waves. It may be more reasonable to use beta waves as an analytical indicator than alpha waves for identifying the EO state.
Most of the researches [
In our case the data processing includes data import, normalization, segmentation, feature extraction, neural network creation, training and test steps.
To facilitate the activity classification, the raw data was divided into small segments (windows). The main challenge in this task is the find the proper window size. In order to improve the accuracy, we made diferent degrees of linear enhancement according to the correlation of each channel. In case of PSD calculation, multiple FFT window sizes were tested. Windows were overlapping, at each sample the window contained the current sample and the previous N-10 samples. Before classification, a 5-Fold Cross-Validation was used for randomly shufled the training and testing data set.
We tested and compared the performance of feedforward MLP networks and CNN in this step. In case of MLP, training function was Levenberg-Marquardt (trainlm). We achieved the best results with the two hidden layers network contained 12 and 7 neurons using log-sigmoid transfer functions (as shown in Figure 6). Output layer had only one neurons as it is two states (EO and EC) to be recognized. On the layers the initial weights came from a normal Gaussian distribution. In the training algorithm the epoch limit was 1000 cycles.
The advantage of deep learning algorithm classification accuracy is usually relfected when the number of sample sets is large enough. And the more complex the network, the more parameters to be trained, the more training set samples are needed. Therefore, the high complexity of the network model cannot be pursued blindly in the design of neural networks. Layers of the used CNN is shown in Table 1. The raw data were converted to grayscale images and these images were used as inputs for the CNN. Image size was 64 × 1 × 8 (64 measurement points, 8 channels).
In addition to the design of network structure, many hyperparameters still need to be determined manually in the training process of a CNN classifier. We created a set of options for training a network using stochastic gradient descent with momentum.
We used a mini-batch with 64 observations at each iteration. If the batch size is set too small, it will be dificult for the network to converge and under-fitting. If the batch size is set too large, it will result in reduced eficiency or memory overflow.
Learning rate is a very important hyperparameter in network training. On the on hand, if the learning rate is set too small the error curve drops too slowly. On the other hand, too large learning rate will lead to error explosion, and the network cannot find the correct direction of gradient descent. We reduced the learning rate by a factor of 0.1 every 10 epochs with 0.2 initial learning rate and set the maximum number of epochs for training to 50.
We set L2 regularization to 0.0005. It ensures that feature weights are not too large, and the processed feature weights are relatively average.
All the following experimental results are from the same computer(CPU: Intel Core i5-7300HQ, RAM: 16 GB,GPU: GeForce GTX 1050 Ti) with MATLAB R2020a.
Classification model evaluation metric was accuracy, calculated the following way:
Accuracy =
No correct predictions
No all predictions .
To determine a proper window size for band power calculation, diferent window sizes were used, training and testing were repeated 1000 times to decrease statistical uncertainty. Used neural network in this case was the 2-layer MLP with 12-7 neurons in hidden layers. The Mean and best accuracy on test data using diferent windows are summarized in Table 2. In case of real-time data processing, we hope to find a window small enough. We found window size of 40 samples a best choice as we are able to identify the activities of the subjects from the EEG signal in approximately 0.25 second. It is an acceptable latency and gives significantly better accuracy than other windows. Therefore, all of the following experiments are based on this window.
In order to further improve the accuracy of the network, we enlarged the data set. Each segment has 40 points that overlap the previous segment.
The best result obtained after enlargement was 96.63% for test data. The comparison of training time and mean accuracy are showed in Table 3. In Table 3, the sufix with A represents only alpha waves considered as feature. The sufix with B represents only beta waves considered as feature. The sufix with A and B represents the feature combined by alpha waves and beata waves. The sufix with L represents using expanded data set. All the accuracy results are the average values of 100 time 10-fold cross-validation.
In this paper we proposed two artificial neural network approaches for EO and EC tasks recognition from EEG data. We used data from 50 volunteers and tried to recognize their activities with feedforward MLP network and CNN in combination with diferent data processing approaches. Our results demonstrated that the determination of the activity from the EEG signal is possible with high classification accuracy.
In the case of MLP, the results obtained show that the accuracy obtained using only alpha or beta waves are not significantly diferent, but using the extended data set makes a significant diference. So it would be more reasonable to use beta wave as an analytical indicator than alpha wave for this this purpose. This can provide some reference for studying the EEG signal of the subject’s eye state.
We reached the higher accuracy rate and shorter training time of using MLP instead of CNN for this purpose. Compared with MLP, however, CNN combines signal preprocessing, feature extraction and classification. It prevents the blindness and cumbersomeness of EEG signal processing, and it also has a good accuracy rate. It is essential to produce a CNN construction with high robustness performance and accuracy rate for EEG signals.
The data set in this project was subjected to a simple preprocessing (bandpass iflter and the normalized processing) without special processing for ECG, EMG and noise. The data set itself has much room to improve, and then makes the accuracy of the model also have great potential for ascension. Due to the limited research time of the project, more details in parameter adjustment have not been studied. However, the eficiency of deep neural network models such as CNN is largely dependent on parameter adjustment. Therefore, the further improvement of model performance needs to be further improved in parameter adjustment.