Unsupervised Seizure Detection based on Rhythmical Activity and Spike Detection in EEG Signals

In this work an unsupervised epileptic seizure detection methodology is proposed to automatically isolate candidate EEG seizure segments without requiring any apriori information or human intervention. Seizures are detected when a concentration of rhythmical activity is observed in the lower end of the EEG spectrum within the delta (≤3 Hz), theta (4-7 Hz) and alpha (8-13 Hz) frequency bands, along with an increase in EEG signal variation and a high frequency of spikes/sharp waves in the same segment. When a candidate seizure segment is detected, channel localization is also evaluated to assess its activation pattern providing feedback to the clinician regarding potential seizure onset area. The proposed methodology is evaluated using EEG signals from two open EEG datasets: the CHB-MIT Scalp EEG Database with long-term recordings and the TUH Seizure Corpus consisting of selected EEG epochs with high ictal to interictal duration ratio with an average sensitivity of 81.15-91.35% and a false detection rate between 5.33-12.15 FD/h.