Motor Imagery EEG Classification Using Common Spatial Pattern Method Expanded to Multidimensional Features

This study focuses on improving classification accuracy of motor imagery-based EEG signals using a userfriendly, wearable EEG device. Participants were instructed to perform left and right hand motor imagery tasks, during which EEG signals were recorded. To address noise and signal instability caused by loose or poorly contacted electrodes, Independent Component Analysis (ICA) was applied as a preprocessing step to extract reliable components. For feature extraction, we used Common Spatial Patterns (CSP) and extended its feature space from the conventional two dimensions to ten dimensions by selecting multiple spatial filters from both ends of the eigenvector distribution. The logarithm of the variance (logvar) was then computed for each CSP component. For classification, we adopted AdaBoost, an ensemble learning method that combines multiple weak classifier ―in this case, depth-1 decision trees―to improve overall prediction performance. Evaluation was conducted using 5-fold cross-validation on data collected from five participants. Results showed that increasing the number of CSP dimensions led to improved classification accuracy for most subjects. The highest accuracy achieved was 71.3%, compared to 57.8% in our previous method using only 2D CSP and SVM. These findings demonstrate the effectiveness of combining ICA preprocessing, multidimensional CSP feature extraction, and AdaBoost classification in enhancing motor imagery EEG decoding performance.