Epilepsy is a neurological disorder that affects over 2% of the world population. Epilepsy patients suffer from recurring seizures that can be very harmful. The unpredictability of seizures is a major concern for medical practitioners because uncontrollable seizures can lead to sudden death and morbidity. A system that could warn patients and doctors alike about the impending seizure event would dramatically enhance the quality of life for patients.

While most previous research works focused on using signal processing tools appropriate for stationary signals, we propose here to use time and frequency (TF) analysis to extract features capable of discriminating normal from abnormal EEG traces (both ictal and interictal). The features are extracted using Singular Value Decomposition (SVD) of the EEG signal Time Frequency matrix. The left singular vectors of the time frequency matrix are used to obtain robust feature vectors. In contrast to existing techniques, the proposed TF-based technique can be used to detect the specific moments of seizure occurrences in time so that this information is used to discriminate interictal from ictal EEG traces. Instead of extracting the features directly from the TF matrix, we transform the left eigenvectors obtained from the SVD of the TF matrix into a feature vector that behaves like to a probability density function.

We show that almost all classical classification techniques achieve excellent seizure detection results when used with the proposed TF features, irrespective of the classifier used. Contrary to existing works, we test our approach across several real-life scenarios covering 2, 3, and 5 possible classes of data. Our tests provided consistent results across different scenarios. The results, under different scenarios, outperformed existing ones achieving consistently more than 97.3% and up to 99.5% in terms of accuracy, sensitivity, and specificity.

Experimental results show that the novel features have successfully represented the characteristics of the underlying disease phenomenon from EEG data. Also, we conclude that learning based classifiers are better suited for this application, compared to Bayesian classifiers that have difficulty in adapting to the varying nature of the features' probability distribution function.