Neurosurgical therapy of refractory epilepsy requires accurate localization of seizure onset zone (SOZ). In clinical practice, intracranial EEG (iEEG) is recorded in the epilepsy monitoring unit (EMU) over many days where multiple seizures are recorded to provide information to localize the SOZ. The prolonged monitoring in the EMU adds to the risk of complications and can include intracranial bleeding and potentially death. Recently, high frequency oscillations (HFO) of iEEG between 80 to 500 Hz are highly valued as a promising clinical biomarker for epilepsy. HFOs are believed to be clinically significant, and thus could be used for SOZ localization. However, HFOs can also be recorded from normal and non-epileptic cerebral structures. When defined only by rate or frequency, pathological HFOs are indistinguishable from physiological ones, which limit their application in epilepsy pre-surgical planning. In this proposal, to the best of our knowledge, we show of a recurrent waveform pattern that distinguishes pathological HFOs from physiological ones. In particular, we observed that the SOZ generates repeatedly a set of stereotyped HFO waveforms whereas the HFOs from nonepileptic regions were irregular in their waveform morphology. Based on these observations, using computational tools built on recent advances in sparse coding and unsupervised machine learning techniques, we propose to detect these stereotyped recurrent HFO waveform patterns directly from the continuous iEEG data of adult and pediatric patients and test their prognostic value by correlating the spatial distribution of detected events to clinical findings such as SOZ, resection zone and seizure freedom. We hypothesize that accurate detection of pathologic HFOs in brief iEEG recordings can identify the SOZ and eliminate the necessity of prolonged EMU monitoring and reduce the associated risks. With these motivations, in this project an interdisciplinary team composed of biomedical engineers, epileptologists and neurosurgeons will work together to develop and test novel computational tools to detect stereotyped HFOs and its subtypes in large iEEG datasets recorded with clinical electrodes. Developed algorithms and iEEG data will be shared with the research community to contribute to the reproducible research and help other research groups to develop novel methods. The results of this study will be essential for achieving our group's long term goal of developing an online neural signal processing system for the rapid and accurate identification of SOZ with brief invasive recording.
Prolonged iEEG monitoring for SOZ localization does add to the risk of complications and may include serious issues, such as intracranial bleeding, meningoencephalitis, and eventually death. The intellectual merit of this project is to develop computational intelligence tools based on recent advances in sparse coding and unsupervised machine learning techniques to investigate stereotyped high frequency oscillations (HFOs) in long-term iEEG and test the hypothesis whether the automated detection of HFOs will yield accurate and fast identification of SOZ.