Subdural EEG classification into seizure and nonseizure files using neural networks in the gamma frequency band

This study describes a new method for offline seizure detection using intracranial EEG (iEEG). The proposed method integrated two interrelated steps: (1) establishing a decisional space on the basis of the interelectrode mean of the spectral power in the gamma frequencies after a thorough evaluation of temporal and frequency-based features and (2) constructing an artificial neural network that operated on this decisional space to delineate EEG files that contained seizures from those that did not. The data were obtained from 14 patients who underwent two-stage epilepsy surgery with subdural recordings. Of the total 157 files considered, 35 (21 interictal and 14 ictal) iEEG data files or 22% were selected randomly and used initially in a training phase. The remaining 122 iEEG data files or 78% were then used in the testing phase to assess the merits in selecting gamma power as means to detect a seizure. The results obtained exhibited an accuracy of 95.90%, a sensitivity of 92.59%, and a specificity of 96.84%. Although this method had to contend with the complex nature of iEEG and the inherent heavy computational load, the constructed artificial neural networks together with the chosen decisional space yielded the best possible outcome. The proposed method was based on aggregating the power in the 36 to 44-Hz frequency range and analyzing its behavior in time, looking for patterns indicative of seizure evolution. It was shown that the power measurement in the gamma range contains the information needed to discriminate seizure files from nonseizure files. The algorithm consisted in establishing a decision space most suitable for iEEG data classification by relying on the power spectra in the gamma frequencies and constructing and implementing an artificial neural network that generates the highest classification accuracy possible. It was noted that although only 29% (35/122) of the files were used randomly for training the detector, high measures in sensitivity, specificity, and accuracy were still achieved in the remaining files, which were subsequently used in the testing phase. Seizures are known to occur intermittently and unpredictably, and massive amounts of EEG or iEEG data need to be analyzed offline to detect seizures. This is a challenge that can only be met through reliable and time-efficient seizure-detection paradigms, an affirmation this study attempted to prove.