Discerning and quantifying high frequency activities in EEG under normal and epileptic conditions

TL;DR

This study analyzes high-frequency EEG spectra to distinguish between normal and epileptic brain activity, revealing specific spectral and dynamic differences that enable high-accuracy classification of epileptic states.

Contribution

The paper introduces a novel approach combining Fourier reconstruction, Welch's transform, and phase-space analysis to identify and quantify high-frequency EEG features associated with epilepsy.

Findings

Identified significant differences in gamma band (40-100 Hz) between normal and epileptic EEGs.

Achieved 94-95% accuracy in classifying epileptic versus healthy states using machine learning.

Revealed bi-stability and bifurcations in EEG dynamics related to epileptic episodes.

Abstract

We investigate the nature of the modifications in the temporal dynamics manifested in the high-frequency EEG spectra of the normal human brain in comparison to the diseased brain undergoing epilepsy. For this purpose, the Fourier reconstruction is efficaciously made use of after Welch's transform, which helped identify the relevant frequency components undergoing significant changes in the case of epilepsy. The temporal dynamics involved in the EEG signals and their associated variations showed a well-structured periodic pattern characterized by bi-stability and significant quantifiable structural changes during epileptic episodes. In particular, we demonstrate and quantify the precise differences in the high-frequency gamma band (40-100 Hz) present in EEG recordings from neurologically normal participants compared to those with epilepsy. The periodic modulations at two dominant frequencies around 50 Hz and 76 Hz in power spectral density are isolated from high frequency noise through the use of Welch's transform, pinpointing their collective behaviors through a phase-space approach. The reconstructed signals from these restricted frequency domains revealed oscillatory motions showing a bi-stability and bi-furcations with distinct differences between normal and seizure conditions. These differences in the phase space images, when analyzed through linear regression and SVM-based machine learning models, support a classification accuracy of around 94-95% between healthy and ictal states using a publicly available EEG dataset from the University of Bonn (Germany). The partial reconstruction of the dynamics as compared to the earlier studies of the full phase space accurately pinpointed the destabilization of the collective high-frequency synchronous behavior and their precise differences in the normal and diseased conditions, avoiding the other chaotic components of the EEG signals.