Abstracts

Rationale: Manual Epileptic Seizure Prediction from raw Electroencephalogram (EEG) data is a cumbersome activity in Clinical Neuroscience. Computational tools like ‘Support Vector Machines’ (SVMs) and ‘Artificial Neural Networks’ can be exploited to automate this process, by modeling Epileptic Seizure Prediction as a classification problem. The application of an SVM in Epileptic Seizure Prediction is discussed. An SVM is a supervised learning algorithm which addresses the general problem of learning to discriminate between ‘positive’ and ‘negative’ members (which constitute the “Training Set”) of a given class of n-dimensional vectors. The “Training Set” is fed as input to the SVM to ‘train’ the SVM i.e. to enable it to define a separating hyperplane, which separates the positive and negative samples, which in this case are the EEG recordings that denote normalcy and seizure activity, respectively. Signal analysis aims to extract appropriate information from a signal. The notion of time-scale signal analysis using the Discrete Wavelet Transform (DWT) is recognized as a potential tool for analyzing EEG signals. The DWT representation of a signal is obtained by passing the original signal through two complementary filters that emerge as two signals (digital filtering techniques). The signal passed through the low-pass filter includes the high-scale, low-frequency components (called ‘approximation’) and that passed through the high-pass filter includes low-scale, high-frequency components (called ‘detail’). Methods: For this study, the EEG readings were taken from http://www.epileptologie-bonn.de/front_content.php?idcat=193&lang=3&changelang=3. The sampling rate of the data is 173.61 Hz. The time series have the spectral bandwidth of 0.5 Hz to 85 Hz of the acquisition system. Firstly, a low-pass filter of 40 Hz has to be applied on the data. Wavelet transform using Daubechies 4th order Mother wavelet function “db4' is employed for decomposing the EEG signal, since it gives sufficient resolution and is proved to be efficient for seizure detection. A new set of wavelet features viz. Mean, Mean of absolute values, Average Power, Standard Deviation and Cross Correlation are extracted. These five parameters constitute one input vector for the SVM. The following algorithm will enable the use of SVM for Epileptic Seizure Prediction: 1. Standard EEG recordings (with and without Seizure activity) are taken. 2. For each frame of EEG (corresponding to 1 second recording of EEG signals), the wavelet features aforementioned, are extracted. 3. Once all frames are processed, readings that are normal and those with seizure activity are labeled +1 and -1 respectively. 4. The SVM is ‘trained’ with the training set and made to find the decision boundary. The SVM is now ready for Epileptic Seizure Prediction of any new EEG recording. Results: Sets of manually pre-classified EEG signals were fed as test-inputs into the SVM; it was found that the SVM was 90.1-94.7% accurate in Predicting Seizure activity. Conclusions: Thus, employing an SVM is an efficient solution for automated Epileptic Seizure Prediction.