Investigating the effects of Gaussian noise on epileptic seizure detection: The role of spectral flatness, bandwidth, and entropy

Abstract

This study investigates the effect of Gaussian noise on the classification of EEG signals from five classes in the Bonn University EEG dataset for epileptic seizure detection, using Power Spectral Density features. The EEG data are pre-processed with a low-pass filter at a cutoff frequency of 40 Hz, and a total of 11 features, including spectral flatness difference, spectral bandwidth difference, and entropy difference, are extracted. Feature vectors are generated for both original signals and signals with varying levels of injected Gaussian noise. The results demonstrate that noise injections significantly improve classification accuracy across all class combinations by enhancing feature separability and generalization. Notably, 100 % accuracy was achieved in classifications with different noise levels. Analyses performed using classifiers such as Random Forest, Multilayer Perceptron, and k-Nearest Neighbors show that the Random Forest classifier achieves high classification success across all noise levels. Additionally, it was found that incorporating spectral flatness difference, spectral bandwidth difference, and entropy difference features significantly contributes to classification accuracy when combined with noise injection. This study highlights the potential of noise injections to reduce overfitting and enhance the robustness of EEG classification, providing valuable insights for future biomedical signal analysis. Noise injection, traditionally viewed as a factor that could hinder performance, is utilized in this study as a novel approach to enhance classification accuracy, marking a significant innovation in the field.