COMPARATIVE ANALYSIS OF ML AND DL CLASSIFIERS FOR EEG-BASED EPILEPTIC SEIZURE DETECTION USING PERFORMANCE METRICS

Abstract

Neurological disorder known as epilepsy is typified by aberrant brain neuronal activity that results in frequent, unprovoked seizures.  It represents about 1% of all diseases worldwide and affects more than 70 million people. The mortality rate for people with epilepsy is almost three times that of the total population.  However, research indicates that with the right diagnosis and care, up to 70% of epilepsy cases can be effectively managed and possibly resolved.  Using a variety of machine learning (ML) and deep learning (DL) classifiers, the goal of this review is to suggest appropriate seizure detection methods that are customized for specific medical datasets.  The feature extraction methods employed, such as Multi-Fractal Detrended Fluctuation Analysis (MFDA), Variational Mode Decomposition (VMD), and Discrete Wavelet Transform (DWT), have a major impact on these classifiers' performance. Accurate seizure detection depends on choosing the best feature and classifier combination. Machine learning techniques that have achieved 100% accuracy include Random Forest with L1-PRR and LS SVM with Tunable Q-Factor Wavelet Transform (TQWT) with fractal dimension.  An Artificial Neural Network (ANN) that used Local Neighborhood Difference Pattern (LNDP) and 1-D Local Gradient Pattern (LGP) for feature extraction demonstrated 98.96% accuracy, 98.56% sensitivity, and 99.38% specificity, demonstrating the immense promise of deep learning models.  Additionally, 10-fold cross-validation was used to achieve 100% accuracy for ED-LSTM.  Convolutional Neural Networks (CNN) and K-Nearest Neighbours (KNN) are two more classifiers that have demonstrated impressive performance. In order to improve the identification of epileptic seizures, this study highlights the significance of choosing the best EEG-based ML and DL classifiers in addition to suitable feature extraction techniques.