Abstracto

Spatial Filter Optimization Using Gaussian Kernel for Single Electro- Encephalo Gram (EEG) Trial Classification

Aarti Bhalla

Brain Computer Interfaces (BCI) enable people suffering from Amyotrophic Lateral Sclerosis (ALS) or severe paralysis to communicate and control using only brain signals. These signals are captured from EEG, that is relatively cost effective and non invasive. Raw EEG signals are subject to processing and feature extraction for automating the detection of changes in the amplitude of the brain signals due to imagining of a limb motion and to enable classification. The processed signal is analysed using machine learning tool and translated into interpretable form. In this manuscript, we propose an advanced feature extraction method utilizing spatial characteristics (channel information) of an EEG signal. The method maximizes inter class variance and minimizes intra class variance among the optimized spatial features to facilitate classification. Kernel based Fisher’s criterion is used for Spatial filtering which transforms the data into a suitable feature space for classification. This approach overcomes the limitation of the existing Fisher based Common Spatial Patterns and performs better in general. Experimental results on BCI Competition datasets demonstrate the effectiveness of this methodology in terms of accurate classification, in comparison to the other method.