Abstract

The success of an epilepsy treatment, such as resective surgery, relies heavily on the accurate identification and localization of the brain regions involved in epilepsy for which patients undergo continuous intracranial electroencephalogram (EEG) monitoring. The prolonged EEG recordings are screened for two main biomarkers of epilepsy: seizures and interictal spikes. Visual screening and quantitation of these two biomarkers in voluminous EEG recordings is highly subjective, labor-intensive, tiresome and expensive. This thesis focuses on developing new techniques to detect and classify these events in the EEG to aid the review of prolonged intracranial EEG recordings.

It has been observed in the literature that reliable seizure detection can be made by quantifying the evolution of seizure EEG waveforms. This thesis presents three new computationally simple non-patient-specific (NPS) seizure detection systems that quantify the temporal evolution of seizure EEG. The first method is based on the frequency-weighted-energy, the second method on quantifying the EEG waveform sharpness, while the third method mimics EEG experts. The performance of these new methods is compared with that of three state-of-the-art NPS seizure detection systems. The results show that the proposed systems outperform these state-of-the-art systems.

Epilepsy therapies are individualized for numerous reasons, and patient-specific (PS) seizure detection techniques are needed not only in the pre-surgical evaluation of prolonged EEG recordings, but also in the emerging neuro-responsive therapies. This thesis proposes a new model-based PS seizure detection system that requires only the knowledge of a template seizure pattern to derive the seizure model consisting of a set of basis functions necessary to utilize the statistically optimal null filters (SONF) for the detection of the subsequent seizures. The results of the performance evaluation show that the proposed system provides improved results compared to the clinically-used PS system.

Quantitative analysis of the second biomarker, interictal spikes, may help in the understanding of epileptogenesis, and to identify new epileptic biomarkers and new therapies. However, such an analysis is still done manually in most of the epilepsy centers. This thesis presents an unsupervised spike sorting system that does not require a priori knowledge of the complete spike data.