The Implementation of Feature Selection Techniques in Search for Diagnostic Criteria for Epilepsy

Abstract

In this study, we consider possible application of feature selection methods implemented in the Weka software package for the diagnosis of epilepsy. We analyze bioelectric brain signals for solving the problem of classifying electroencephalograms obtained from two groups of people: patients with epilepsy and healthy subjects. Features were extracted from electroencephalograms using Python framework EEGlib, which is compatible with Weka package. During the analysis, we considered 11 machine learning algorithms included in the Weka package, for which the mean value of classification F score exceeded 0.8 and mean value of AUC ROC reached 0.89 on the original feature space (28 features for classification). Analysis was performed in two stages. At the first stage, we analyzed signals recordings under study using classifying machine learning methods within the initial feature space. After that, we used two evaluators: CfsSubsetEval and WrapperSubsetEval and chose new subset of features of each evaluator based on individual predictive power of features. Then, we trained classifying algorithms on obtained subsets of features for evaluating their effectiveness for solving classification problem. Feature evaluators increased the average F score and AUC ROC of most methods but decreased for those with high values. Best result was obtained for Random Forest algorithm without feature selection and the F score, and AUC ROC of this approach was 0.882 and 0.965 correspondingly.