Date Approved

12-2014

Graduate Degree Type

Thesis

Degree Name

Engineering (M.S.E.)

Degree Program

School of Engineering

First Advisor

Dr. Samhita Rhodes

Second Advisor

Dr. Paul Fishback

Third Advisor

Dr. Robert Bossemeyer

Fourth Advisor

Dr. Kost Elisevich

Abstract

Epileptic seizures are characterized by abnormal electrical activity occurring in the brain. EEG records the seizures demonstrating changes in signal morphology. These signal characteristics, however, differ between patients as well as between different seizures in the same patient. Epilepsy is managed with anti-epileptic medications but in some extreme cases surgery might be necessary. Non-invasive surface electrode EEG measurement gives an estimate of the seizure onset but more invasive intra-cranial electrocorticogram (ECoG) are required at times for precise localization of the epileptogenic zone.

The epileptogenic zone can be described as the cortical area targeted for resection to render the patient symptom free. Epileptologists use the “evolution” of aberrant signals for identifying epileptic seizures and the epileptogenic zone is identified by concentrating on the area contributing to the onset of seizure. This process is done by visually analyzing hours of ECoG data. The signal morphology during an epileptic seizure is not very different from abnormal discharges noticed in ECoG data thereby complicating signal analysis for the epileptologists. This thesis aims to classify the ECoG channel data as epileptic or non-epileptic using an automated machine learning algorithm called support vector machines (SVM). The data will be decomposed into various frequency bands identified by wavelet transform and will span the range of 0-30Hz. Statistical measures will be applied to these frequency bands to identify features that will subsequently be used to train SVM. This thesis will further investigate feature reduction using multivariate analysis methods to train the SVM and compare it to the performance of classification when all the features were used to train SVM.

Results show that channel data classification using trained SVM that did not undergo feature reduction performed better with 98% sensitivity but needed more runtime than the SVM algorithms that was trained using reduced features. For high frequency analysis of frequencies between 60-500Hz, the results show the same sensitivity yet less specificity when compared to the classification using lower frequency range of 0-30Hz.

The results seen in this thesis show that support vector machines classifiers can be trained to classify the data as epileptic or non-epileptic with good accuracy. Even though training the classifiers took almost two hours, it was still noticeably less than other machine learning algorithms such as artificial neural networks. The accuracy of this algorithm can be improved with changes to the data segment length, size of training matrix, accuracy of epileptic and nonepileptic data, and amount of data used for training.

Included in

Engineering Commons

Share

COinS