A versatile EEG spike detector with multivariate matrix of features based on the linear discriminant analysis, combined wavelets, and descriptors

Abstract

The wavelet transform has been used together with many types of classifiers for processing and detecting epilepsy patterns (spikes) in electroencephalographic signals (EEG) in the last 2 decades. A new improved detector architecture is proposed that applies a combination of wavelets and descriptors in a multivariate matrix of features to extract and enhance discriminatory information of spikes. The number of extracted features is reduced by linear discriminant analysis (LDA) that always determines a one-dimensional matrix containing distributions of spike and non-spike samples. A simple linear classifier is used after LDA for binary classification. The area under the curve (AUC) drawn by the receiver operating characteristics (ROC) is used as index to compare performance among different classifier configurations. The result is a classifier architecture that has an AUC index of 0.9941 representing sensitivity and specificity of 97.37% and 97.21%, respectively. The proposed architecture allows different configurations to be tested without changing the classifier architecture and its training is done without iteration.