SVM CLASSIFIER FOR EPILEPTIC SEIZURE PREDICTION USING SUB-BANDING OF IEEG SIGNALS
Epileptic seizure in the brain affects the day-to-day life of any individual due to its unexpected nature of occurrence. It has affected more than 50 million people worldwide. Drug resistance of patients is an important factor which leads to failure of epilepsy treatments using medications in 30% of patients. Surgery is also not a viable option in a substantial number of patients. In such cases, a new kind of seizure forecasting system is necessary to help those people. In our work, various sub-frequency bands of EEG signals are produced from the originally recorded Intracranial Electroencephalogram (IEEG) signals of five canines and two persons to identify possible low complex and less intense EEG features from each sub-band of the entire spectrum. Support Vector Machine (SVM) with different Kernel-based classifiers are used to categorize features into preictal and interictal data. Epileptic Seizures forecasting accuracy of 99% has been achieved for data from canine and human. Employed wavelet filter for noise removal and found that it improved the seizure prediction accuracy in some subjects and reduced the accuracy in some subjects. Similarly, the feature selection technique also improved the preictal detection accuracy in some patients/subjects and reduced the accuracy in some data. From this work, we identified that seizure prediction is possible in at least one sub-frequency band especially in high gamma sub-band generated from the originally recorded signal using a high-pass filter. This work demonstrates an algorithm for seizure forecasting or identifying the preictal region which identifies the suitable best sub-frequency band for predicting the seizure of the originally recorded EEG data by using the computationally less intense EEG features and employing the best classifying SVM kernel.