Detecting Epileptic Seizures from iEEG with Spectral Envelope Analysis and Deep CNNs

Epilepsy is a common chronic neurological disorder characterized by recurrent unprovoked seizures. Electroencephalogram (EEG) and intracranial EEG (iEEG) recordings can provide essential clinical information for epilepsy diagnosis, seizure detection and classification. For seizure detection, the analysis of EEG/iEEG data in the frequency domain is one of the most widely adopted methods. Compared to previous works that focused on the analysis of single-channel EEG/iEEG signals, this paper introduces a multivariate spectral analysis approach, namely spectral envelope, to reveal the frequency patterns that contribute the most to seizure identification. The spectral envelope systematically selects the spectral components that appear simultaneously in multi-channel EEG/iEEG signals, thus enhancing the common periodic signatures. By computing the time-varying spectral envelopes, the dynamic spectral characteristics of multi-channel 1-d EEG/iEEG time series are represented by 2-d time-frequency diagrams, which serve as the inputs to the subsequent classifier. A deep CNN, VGG-16, which has been pre-trained on ImageNet image dataset, is employed. After slight modification in the structure, the VGG-16 network is transferred to the classification task. The above approach was evaluated by experiments conducted on the HUP dataset, a multi-channel iEEG dataset. Compared to single-channel iEEG signals analysis methods including short-time Fourier transform and continuous wavelet transform, the proposed approach better captures the spectral features of multi-channel iEEG signals. The results show the reliability of our method, with an average accuracy, precision, and sensitivity of 98.49%, 98.7%, and 97.91%, respectively.