SVM-Based Pre-Ictal Phase Detection for Epileptic Seizure Analysis

An epileptic seizure is triggered by a sudden and uncontrolled surge of electrical activity in the brain, leading to a range of symptoms like abnormal movements, unusual sensations, loss of consciousness, or, in severe cases, even death. Frequent seizures are the hallmark of the neurological condition known as epilepsy. Seizures can vary in duration, intensity, and manifestation, and can arise from various factors, which encompass genetic susceptibility, brain trauma, infections, or specific medical disorders. Treatment for epilepsy usually involves medications to control seizures, although in some cases, surgery or other interventions may be considered. Pre-ictal phase is signal prior to seizure onset which includes variations in specific features of the signal compare to normal state for seizure prediction. Detection of pre-ictal phase helps to forecast the seizure and helps to save the life with proper treatment. This paper includes extraction of features like peak signal, energy of the signal, standard deviation, derivative of spectrum from pre-ictal phase using 30 minutes and 60 minutes duration before seizure onset. A 7 level DWT decomposition with db6 mother wavelet is used for EEG rhythm extraction. K-nearest neighbor and support vector machine are used to classify the pre-ictal signal. CHB-MIT Scalp EEG dataset of six patients were analyzed by varying the duration of pre-ictal phase for training the classifiers. Misclassification rate, sensitivity and accuracy are used as a routine evaluation parameter. From the result analysis it can be concluded that SVM with linear kernel gives better results with duration of preictal phase of 60 minutes over the test datasets with average accuracy of 97.61%, sensitivity of 97.6 % and misclassification rate of 0.02. To reduce the feature matrix individual band analysis is done which improves the results using features from 30 minutes of pre-ictal phase of theta band with average accuracy of 99.88%, sensitivity of 99.77 % and misclassification rate of zero value.