EPILEPTIC EEG SIGNALS RHYTHMS ANALYSIS IN THE DETECTION OF FOCAL AND NON-FOCAL SEIZURES BASED ON OPTIMISED MACHINE LEARNING AND DEEP NEURAL NETWORK ARCHITECTURE
No Thumbnail Available
Date
2023-06
Journal Title
Journal ISSN
Volume Title
Publisher
World Scientific
Abstract
Objective: Most studies in epileptic seizure detection and classi¯cation focused on classifying
di®erent types of epileptic seizures. However, localization of the epileptogenic zone in epilepsy
patient brain's is paramount to assist the doctor in locating a focal region in patients screened
for surgery. Therefore, this paper proposed robust models for the localization of epileptogenic
areas for the success of epilepsy surgery. Method: Advanced feature extraction techniques were
proposed as e®ective feature extraction techniques based on Electroencephalogram (EEG)
rhythms extracted from Fourier Basel Series Expansion Multivariate Empirical Wavelet
Transform (FBSE-MEWT). The proposed extracted EEG rhythms of ; ; ; and features
were used to obtain a joint instantaneous frequency and amplitude components using a subband
alignment approach. The features are used in Sparse Autoencoder (SAE), Deep Belief
Network (DBN), and Support Vector Machine (SVM) with the optimized capability to develop
three new models: 1. FMEWT–SVM 2. FMEWT SAE–SVM, and 3. FMEWT–DBN–SVM. The
EEG signal was preprocessed using a proposed Multiscale Principal Component Analysis
(mPCA) to denoise the noise embedded in the signal. Main results: The developed models show
a signi¯cant performance improvement, with the SAE–SVM outperforming other proposed
models and some recently reported works in literature with an accuracy of 99.7% using -rhythms in channels 1 and 2. Signi¯cance: This study validates the EEG rhythm as a means of
discriminating the embedded features in epileptic EEG signals to locate the focal and non-focal
regions in the epileptic patient's brain to increase the success of the surgery and reduce
computational cost.
Description
Keywords
EEG; time–frequency; SVM; SAE; DBN; MEWT; FBSE