Electroencephalogram (EEG) pattern recognition problem is considered as a composite of three subproblems: feature extraction, feature selection, and pattern classification. Focusing particularly on the feature selection issue, each subproblem is reviewed briefly and a new method for feature selection is proposed. The method suggests that first one shall extract as much information (features) as conveniently possible in several pattern information domains and then apply the proposed unbiased successive feature elimination process to remove redundant and poor features. From this set select a significantly smaller, yet useful, feature subset that enhances the performance of the classifier. The successive feature elimination process is formally described. The method is successfully applied to an EEG signal classification problem. The features selected by the algorithm are used to classify three signal classes. The classes identified were eye artifacts, muscle artifacts, and clean (subject in stationary state). Two hundred samples for each of the three classes were selected and the data set was arbitrarily divided into two subsets: design subset, and testing subset. A proximity index classifier using Mahalanobis distance as the proximity criterion was developed using the smaller feature subset. The system was trained on the design set. The recognition performance on the design set was 92.33%. The recognition performance on the testing set was 88.67% by successfully identifying the samples in eye-blinks, muscle response, and clean classes, respectively, with 80%, 97%, and 89%. This performance is very encouraging. In addition, the method is computationally inexpensive and particularly useful for large data set problems. The method further reduces the need for a careful feature determination problem that a system designer usually encounters during the initial design phase of a pattern classifier.
|