Dimensionality Reduction Effect Analysis of EEG Signals in Cross-Correlation Classifiers Performance

Abstract

In this paper, it is reported a study conducted to verify whether the dimensionality reduction of electroencephalogram (EEG) segments can affect the application performance of machine learning (ML) methods. An experimental evaluation was performed in a set of 200 EEG segments, in which the piecewise aggregate approximation (PAA) method was applied for 25 %, 50 %, and 75 % settings of the original EEG segment length, generating three databases. Afterwards, cross-correlation (CC) method was applied in these databases in order to extract features. Subsequently, classifiers were built using J48, 1NN, and BP-MLP algorithms. These classifiers were evaluated by confusion matrix method. The evaluation found that the reduction of EEG segment length can increase or maintain performance of ML methods, compared to classifiers built from EEG segments with original length in order to differentiate normal signals from seizures.