Abstracts

Rationale: Study the suitability of support vector machine (SVM) algorithms for early seizure detection in animal models of temporal lobe epilepsy. Methods: SVMs are a general class of machine learning algorithms and have been proposed for seizure detection previously. A major issue with using SVMs for seizure detection is that the amount of ictal data, if it is available, is usually minuscule compared to the amount of inter-ictal data. Gardner et al (J of Machine Learning Research, 7:1025-1044, 2006) showed that seizure detection can be viewed as a novelty detection problem and proposed the usage of a method called 1-Class SVM which requires only inter-ictal data for its training. A significant shortcoming of their approach is that in cases where ictal data is available it is completely ignored, as it cannot be incorporated in the method. Support Vector Data Description (SVDD) is a related technique where, unlike 1-Class SVMs, ictal data when available can be used to make the detector more sensitive to seizures. Here we compare the performance of 1-Class SVM with that of SVDD for seizure detection and report the relative advantages of each method. Results: Four EEG signal features - signal energy, mean curve length and wavelet power - were extracted from the data that has both ictal and inter-ictal periods. The detection performances of 1-Class SVM and SVDD in terms of sensitivity, specificity and detection delay are compared using results of 2 fold cross validation on the extracted features. It was found that SVDD when trained on both inter-ictal and ictal data outperforms 1-Class SVM by giving almost 22% lower detection delay. But the specificity of SVDD was 9% lower than that of 1-Class SVM. Both algorithms gave 100% sensitivity. Conclusions: Our experimental results suggest that when ictal data is also available SVDD is more efficient for seizure detection than 1-Class SVMs.