A robust unsupervised epileptic seizure detection methodology to accelerate large EEG database evaluation

In this work an unsupervised methodology for the detection of epileptic seizures in long-term EEG recordings is presented. The design of the methodology exploits the available medical knowledge to tackle the lack of training data using a simple rule-based seizure detection logic, avoiding complex decision making systems, training and empirical thresholds. The Short-Time Fourier Transform is initially applied to extract the EEG signal energy distribution over the delta (<4 Hz), theta (4-7 Hz) and alpha (8-13 Hz) frequency bands. A set of four novel seizure detection conditions is proposed to isolate EEG segments with increased potential of containing ictal activity, by identifying segments where the EEG signal energy is intensively accumulated among the three fundamental frequency rhythms. A set of candidate seizure segments is extracted based on the intensity of the accumulated EEG activity per seizure detection condition. The clinician has to visually inspect only the extracted segments instead of the entire duration of the patient's EEG recordings to speed up the annotation process. The results from the evaluation with 24 cases of long-term EEG recordings, suggest that the proposed methodology can reach on average up to 89% of seizure detection sensitivity, by automatically rejecting 95% of the total patient's EEG recordings as non-ictal, without requiring any apriori data knowledge.