IMPROVING THE ACCURACY OF MOTOR IMAGERY-BASED EEG SIGNAL CLASSIFICATION BY MEANS OF THE IDENTIFICATION AND ELIMINATION OF OUTLIERS

• Main Author: Anggraini M. L. Tobing, Tabita
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/52306
• Institution: Institut Teknologi Bandung

Motor imagery-based EEG signal faces a problem with accuracy for more than two-classes classification. By applying the power spectrum, the beginning idea of classification features unit, it is difficult to obtain a classification accuracy of more than 70%. Previous studies have examined the increase in classification accuracy based on the EEG acquisition system, noise identification and elimination, and the combination of elements in the classification model in the brain-computer interface (BCI) system. Therefore, through this research, a method of outlier identification and elimination is proposed to improve the accuracy of the EEG signals based on three classes of motor imagery classification. The method is statistical-based, namely boxplot. This method analyzes the distribution of data using the interquartile range calculation, the range between the data values in the 25th percentile (Q1) and the data values in the 75th percentile (Q3). This method serves to determine the maximum and minimum limits of the sample for each classification feature and to eliminate samples that are outside these limits. This method is then grouped based on the number of eliminated outliers or the percentage outliers of the total data that are 10%, 20%, 30%, and all. This study processes data sourced from the EEG Motor Movement/ Imagery Dataset managed by a complex physiological signal research agency in America, with open access via www.physionet.org. The classification model includes feature extraction, feature selection, feature dimension reduction, evaluation of the classification model, and classification using machine learning. The number of features available is 56 features that are maximum value power using the fast Fourier transform (FFT) and welch transformation, absolute value power v the average power value, the variance power, skewness power value, and kurtosis power value. The power signal comes from the wave frequency range ? from the O1 and O2 electrodes, the ? wave from the C3, CZ, and C4 electrodes, and the ? wave from the electrodes C3, CZ, and C4. The identification and elimination methods themselves applied after feature selection elements. The increase in classification accuracy is based on the comparison of the classification metric value and the best data testing score category between before and after the implementation of outlier elimination. The outlier identification and elimination algorithms proposed can increase the classification score for data testing both collectively and individually and based on the shift in the trend of the best testing data score category. The highest value of testing data prediction accuracy from one of the subjects is 90% with an increase of 11.05%.