This item is available under a Creative Commons License for non-commercial use only
Epilepsy is one of the most common neurological disorders, and aﬄicts approximately 70 million people globally. 30-40% of patients have refractory epilepsy, where seizures cannot be controlled by anti-epileptic medication, and surgery is neither appropriate, nor available. The unpredictable nature of epileptic seizures is the primary cause of mortality among patients, and leads to signiﬁcant psychosocial disability. If seizures could be predicted in advance, automatic seizure warning systems could transform the lives of millions of people. This study presents a performance comparison of artiﬁcial neural network and sup port vector machine classiﬁers, using EEG spectral features to predict the onset of epileptic seizures. In addition, the study also examines the inﬂuence of EEG window size, feature selection, and data sampling on classiﬁcation performance. A total of 216 generalised models were trained and tested on a public seizure database, which contained over 1300 hours of EEG data from 7 subjects. The results showed that ANN outperform SVM, when using spectral features (p = 0.035). The beta and gamma frequency bands were shown to be the best predictors of seizure onset. No signiﬁcant diﬀerences in performance were determined for the dif ferent window sizes, or for the feature selection methods. The data sampling method signiﬁcantly inﬂuenced the performance (p < 0.001), and highlighted the importance of treating class imbalance in EEG datasets.
Tennant Watson, Ian Thomas (2018). A performance comparison of neural network and SVM classifiers using EEG spectral features to predict epileptic seizures. Masters dissertation, DIT, 2018.