Classification of epileptic seizure using hybrid deep learning framework with time and time-frequency Hjorth features

In this work, a novel framework is proposed which includes Hjorth parameters as features from time and time-frequency domain (Multi-Domain) and attention-enhanced temporal modeling, to classify epileptic seizure stages, namely normal, inter-ictal, and ictal. Three different approaches are compared, i.e. Hjorth parameters in time domain, time-frequency domain, and multi-domain. In time-frequency domain, Hjorth parameters are derived from the wavelet coefficients obtained using Discrete Wavelet Transform (DWT). The extracted features are then fed to a 1D Convolutional Neural Network (CNN), Bidirectional Long Short-Term Memory (BiLSTM), and attention mechanism. The performance of the proposed framework is evaluated on Bonn EEG dataset using different performance evaluation metrics namely precision, recall, F1-score, and accuracy. The binary, three-class, and five-class seizure classification are examined using the proposed framework. The validation of the model is performed through the 10-fold cross-validation with sample level partitioning. Experimental findings show that the proposed framework with multi-domain features has given outstanding performance with 98.40, 98.00, and 85.40% test classification accuracy for binary, three-class, and five-class discrimination, respectively.