Study and comparison of 2 class classification problem in Brain computer interface
Motor Imagery Signal Classification using EEG and ECoG signal for Brain Computer Interface
Md. Redwan Islam Student no-1006066, Umme Fatema Student no 1006194,
Supervised by: Dr. Mohammed Imamul Hassan Bhuiyan, Professor,Department of electrical & electronic engineering,Bangladesh university of engineering & technology.
IntroductionBrain Computer interfacing(BCI) is a communication system where commands or messages are sent to a device or computer by means of ones mind or brain activity.BCI is an emerging technology dealing with computer-aided control using exclusively brain activity, and has found application across bio-engineering fields and in neuroprosthetics.
Importance of BCICommunication through mind is an advanced and simpler way of interaction for fully paralyzed or partially paralyzed patients communication through mind is easier or only possible way.Improves the life style.Gives severely paralyzed people another way to communicate, a way that does not depend on muscle control
Data acquisition for BCI
ECoG (Electrocorticography)Partially non-Invasive BCI data acquisition technique. Electrodes placed directly on the exposed surface of the brain to record electrical activity from the cerebral cortex.
EEG (Electroencephalography)Noninvasive BCI data acquisition methodElectrodes placed along the scalp, although invasive electrodes are sometimes used in specific applications.
ObjectiveTo develop a method for 2 class classification of partially invasive BCI using ECoG signal. To develop a method for 2 class classification of non-invasive BCI using EEG signal.To compare the classification accuracy of the developed methods with other previously proposed methods using Brain Computer Interface BCI Competition databases.
General scheme of BCI
Classification of ECoG signalSince placed directly on the exposed surface of the brain, ECoG signals are obtained at a lower noise and higher sampling rate. Typical sampling rate is 1kHz. We are using BCI Competition III dataset I. The dataset contains ECoG signals corresponding to imagined left pinky and tongue movement.
Database I:BCI competition III dataset IThe subject was not a lock-In patient but suffered fromepilepsy. Recording consisted of 278 trails each either left pinky or tongue movement. 88 ECoG platinum electrode grid which was placed on thecontralateral (right) motor cortex.Covered right motor cortex completely and due to its size(approx. 88cm), also partly covered surrounding cortexareas. Sampling frequency was 1000Hz.To avoid VEP (visually evoked potential), the recording intervals started after end of 0.5s visual cue.
Schematics for Classification of imagined left pinky vs tongue from ECoG signal (method 1)
Average p value = 0.000356
Classifiers used for classification Discriminant classifierKNN classifierSVM classifier
ResultsPer sample validation accuracy was calculated. The Highest Accuracy was obtained by using SVM classifier with radial bias kernel and with a spread of 16 and it was about 87.6%.
Classification Accuracy with variation of distance metric and number of neighbors in KNN classification
Classification Accuracy with variation of radial bias in SVM classifier
Comparison with other methods using BCI Competition III dataset IResearcher & Publication Proposed methodReported accuracy (%)Hai-bin Zhao et.al. [Classifying ECoG Signals Using Probabilistic Neural Network] [Information Engineering (ICIE), 2010 WASE International Conference on (Volume:1 )]Common Average Filter Downsample Channel selection by observation Bandpower Windsorizing PNN85.6
Onder AYDEMIR et. Al. [Wavelet Transform Based Classification of Invasive Brain Computer Interface Data][RADIOENGINEERING, VOL. 20, NO. 1, APRIL 2011 ] Wavelet coefficient mean & variance SVM classifier86.1
We obtained higher accuracy and using further feature selection algorithm, feature length can also be reduced for faster classification.
Classification of EEG signalSince placed on the scalp, EEG signals are obtained at a with noise due to probe movement and neural activity of the skin. Typical sampling rate is 100Hz. We are using BCI Competition IV dataset I which contains imagined left hand and right hand movement.
Database II:BCI Competition IV Dataset IRecorded from a healthy person. Cues were displayed for a period of 4s during which the subject was instructed to perform the cued motor imagery task.
Schematics for Classification of imagined left hand and right hand movement from EEG signal (method 2)
Cepstrum Analysis. Cepstrum analysis is a nonlinear signal processing technique which is frequently used in speech processing. Thecepstrumof a sequencexis calculated by finding the logarithm of the magnitude of the Fourier transform ofx, then the inverse Fourier transform of the resulting sequence.
Common Spatial PatternsCommon spatial pattern (CSP) is a mathematical procedure which determines spatial filtersthat maximize the variance of signals in one class and simultaneously minimize the variance of signals in the other class. After obtaining the transformed projection, the data matrix is classified using the mentioned classifiers.
Average P values = 0.0004
Results obtained from different classifiers Using KNN classifier, 100% accuracy was obtained. Using SVM classifier, linear kernel gave a lower accuracy, but the Gaussian kernel gave 98.9% accuracy.
Comparison with other methods using BCI Competition IV dataset IResearcher & Publication Proposed methodReported accuracy (%)Cheolsoo Park et.al. [Classification of Motor Imagery BCI UsingMultivariate Empirical Mode Decomposition] [IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 21, NO. 1, JANUARY 2013]MEMD Selection of IMF CSP classification 85.6
We obtained about 98.9% success. Besides we used cepstrum analysis which avoids splitting of signals into multiple signals and only changes the domain of the signal which is also suitable for real time BCI.
ClassificationAfter obtaining the final data matrix, we use this data along with their labels to train a statistical classifier so that it can predict the unlabeled data. Untrained modelTrained Model
Training Data Labels Trained Model
Test Data Predicted Labels
Discriminant ClassifierThe idea of Linear Discriminant Analysis (LDA) is to find a projection where class separation is maximized.The goal of LDA is to give a large separation of the class means while also keeping the in-class variance small. Using the kernel trick, LDA is implicitly performed in a new feature space, which allows non-linear mappings to be learned.
KNN classifierKNN is a simple algorithm that stores all available cases and classifies new cases based on a similarity measure.
Support Vector Machine classifierAn SVM classifies data by finding the best hyperplane that separates all data points of one class from those of the other class. Thebesthyperplane for an SVM means the one with the largestmarginbetween the two classes. Margin means the maximal width of the slab parallel to the hyperplane that has no interior data points.Some binary classification problems do not have a simple hyperplane as a useful separating criterion. For those problems, there is a variant of the mathematical approach that retains nearly all the simplicity of an SVM separating hyperplane. This approach uses these results from the theory of reproducing kernels
Accuracy and performance measurement of the methodWe calculated the accuracy of our method by calculating the per sample validation accuracy for the available data. Performance was measured by varying the number of data and calculating the required time in a specific environment. In our analysis on BCI Competition III dataset I, we used MATLAB 2015 software for our analysis in Windows 10 platform with Intel Core i5 processor. The Highest Accuracy was obtained by using SVM classifier with radial bias kernel and with a spread of 12 and it was about 86.6%.
Sequential feature selectionIt is the process that selects a subset of features from the data matrix X that best predict the data in y by sequentially selecting features until there is no improvement in prediction. Rows of X correspond to observations; columns correspond to variables or features. Statistical forward feature selection greatly reduces feature length. Thus the size of the feature matrix reduces to N*M; where M < M. M is the number of features selected after feature selection.