Reducing Electrodes based on Decision Tree Classification for EEG Motor Movement Data

Analysis of EEG data is one of the most important parts of Brain Computer Interface systems because EEG data consists of a substantial amount of crucial information that can be used for better study and improvements in BCI system. One of the problems with the analysis of EEG is the large amount of data that is produced, some of which might not be useful for the analysis. Therefore identifying the relevant data from the large amount of EEG data is important for better analysis. The objective of this study is to find out the performance of Random Forest classifier on the motor movement EEG data and reducing the number of electrodes that are considered in the EEG recording and analysis so that the amount of data that is produced through EEG recording is reduced and only relevant electrodes are considered in the analysis. The dataset used in the study is Physionet motor movement/imagery data which consists of EEG recordings obtained using 64 electrodes. These 64 electrodes were ranked based on their information gain with respect to the class using Info Gain attribute selection algorithm. The electrodes were then divided into 4 lists. List 1 consists of top 18 ranked electrodes and number of electrodes was increased by 15 [in ranked order] in each subsequent list. List 2, 3 and 4 consists of top 33, 48 and 64 electrodes respectively. The accuracy of random forest classifier for each of the list was compared with the accuracy of the classifier for the List 4 which consists of all the 64 electrodes. The additional electrodes in the List 4 were rejected because the accuracy of the classifier was almost same for List 4 and List3. Through this method we were able to reduce the electrodes from 64 to 48 with an average decrease of only 0.9% in the accuracy of the classifier. This reduction in the electrode can substantially reduce the time and effort required for analysis of EEG data.