Development of electroencephalogram based machine learning deep learning models for epileptic seizures classification

Abstract

Epilepsy, a long-term neurological dysfunction, affects people in newlinegeneral across every age group. Epilepsy is classified by one of the following: newlinetwo or more reflex(unprovoked) seizures or a single unprovoked seizure with newlinea newline60% likelihood of recurrence within the next decade. Seizures are newlinecharacterised by sudden episodes of involuntary movements that can be newlinepartial or affect the whole body or abnormal electrical activity in the brain newlinecells, often leading to unconsciousness. The contemporary Machine Learning newline(ML) and Deep Learning(DL) approaches that include Independent newlineComponent Analysis(ICA), Support Vector Machines (SVM), and k Nearest newlineNeighbors (KNN), as well as sophisticated DL techniques like Convolutional newlineNeural Networks (CNN), Recurrent Neural Networks (RNN), and Long Short newlineTerm Memory (LSTM) are extensively used for epileptic seizure detection. newlineMoreover, the limitations of these models were identified by this research are newlinelisted as follows: inadequacy in temporal and spatial patterns, adaptability to newlinevaried datasets, achievement of multiresolution representation, flexibility, and newlinecomputational cost of feature extraction approaches, particularly the limited newlinevalue of kernel approximation functions. The ramifications of suboptimal newlinehyper parameter selection in feature classification models are far-reaching, newlinepotentially compromising the generalizability and high-dimensional feature newlinerepresentation. This research presents three ML-DL models to mitigate the newlineidentified limitations. The H-CFCA model that involves utilization of db4 newlineWavelet-based decomposition (WD) for pre-processing the EEG into five newlinefrequency bands, followed by Singular Value Decomposition (SVD) for newlinedimensionality reduction. newline