Epileptic Seizure Classification And Onset Detection Using Hybrid Time–Frequency Domain Features With Machine Learning Models

Abstract

Clinical management of epileptic seizures requires high-quality classification and prediction tools because of the demanding nature of the situation. Within the framework of this study, it will provide a detailed approach to classifying and onset seizures detection, using the time and frequency domain features extraction methods of EEG signals. In proposed method, 8 salient features are extracted: Instantaneous Phase Shift, Temporal Synchronization Index, Fractal Dimension Dynamics, Nonlinear Energy Operator, Instantaneous Frequency, Multiscale Entropy, Phase-Amplitude Coupling, and Bispectral Index. Three algorithms are applied to analyses the extracted features:  Support Vector Machine (SVM), Random Forest (RF), and  K-Nearest Neighbors (KNN). The experimental results indicate that the SVM models achieve accuracy of 99.3% with sensitivity of 98.0 and specificity of 98.5 in classification and in onset detection 98% accuracy is achieved, thus perform better than other available techniques in the literature. The RF and KNN models also provide the values of competitive performance indicators, which indicate the effectiveness of the chosen features in the classification of the state of seizures and non-seizures. As comparted to all three model SVM model gave better result in both classification and onset detection.  The contribution of this work to the development of automated seizure classification systems is providing a solid framework on the subsequent research and clinical packages.