Implementation of deep neural networks for classifying electroencephalogram signal using fractional S-transform for epileptic seizure detection

摘要

Epilepsy is one of the most common neurological diseases of the human brain. It affects the nervous system of brain which shows the impact on an individual's life because of its repetitious occurrences of seizure. Epileptic detection using automatic learning is essential to reduce the substantial work on reviewing continuous electroencephalogram (EEG) signal in spatial and temporal dimensions. A novel methodology is implemented on EEG signals for the detection of epileptic seizure with the combination of fractional S-transform (FST) and entropies along with deep convolutional neural networks (CNN). The original EEG signals are preprocessed with discrete wavelet transform to generate Daubechies-4 (Db4) wavelets. FST is enacted on every segment of the preprocessed signal for time-frequency representation and the features are obtained through entropies. Afterwards, a 15-layer deep CNN with dropout layer and soft-max is used for classification. The experimental results showed that the singular value decomposition entropy are more stable and deep CNN models always performed better for this entropy. A specificity of 98.70%, sensitivity of 97.71%, and accuracy of 99.70% are achieved for the multichannel segment.