Motor Imagery Based EEG Classification by Using Common Spatial Patterns and Convolutional Neural Networks

Abstract

EEG signal processing has been an important and engaging issue over the last three decades. It has been used in the applications ranging from controlling mobile robots to analyzing sleep stages. Previously it was used in the applications of clinical neurology such as detecting epileptic seizure, finding epileptiform discharges, diagnosis of epilepsy, etc. Convolutional Neural Network (CNN) on the other hand is one of the most popular and successful method that has been broadly utilized in machine learning problems such as pattern recognition, image classification and object detection. The proposed study focuses on maximizing the classification performance by combining two of the most successful methods: CSP (Common Spatial Patterns) and CNN. Three different setups have been established in order to observe the changes in the validation accuracy of the classifier. At first, a CNN (four convolution layers and a fully connected layer) structure is trained by feeding the raw data. Secondly, five different filters are applied to the original signal and their outputs are utilized in the training of a CNN having the same structure. Thirdly, the original signal has been transformed via CSP into another space where its spatial features are observed more clearly and then classified by the CNN. It is observed that the combination of CSP and CNN gives the best performance with 93.75% validation accuracy.