2.4. Feature sets based on power spectral density (PSD)

The power spectral density (PSD) of each data segment was used as a compact representation of the raw data, a standard way to represent a time series in many fields, including epilepsy research and machine learning [7,[17], [18], [19]]. The PSD was computed for the raw ECoG, EEG, and EKG signals using Fast Fourier Transform with five-second, non-overlapping Hanning windows. This generally resulted in a feature vector of 641 length. For ECoG, three feature sets were produced from the PSD to compare predictive performance: (1) The average PSD across all subdural electrodes (ECOG-PSD), (2) the PSD from each individual subdural electrode (ECOG-PSDSingleChannel), where the PSD of each channel comprised its own feature set, and (3) the collected PSDs of each individual channel as one feature vector (ECOG-PSDAllChannels). Similarly, for EEG we created two feature sets: (1) the PSD from each individual subdural electrode (EEG-PSDSingleChannel) where the PSD of each channel comprised its own feature set and (2) the collected PSDs of each individual channel as one feature vector (EEG-PSDAllChannels). For EEG/ECOG-PSDAllChannels, the PSD was computed with a resolution of 2.5 Hz, producing, for each 30-second segment, feature vectors of 1344 length for scalp EEG (64 × 21 channels) and (64 × number of channels a given patient had) for ECoG.