Robust Detection of High-frequency Oscillations in Noisy Intracranial EEG Recordings for Early Identification of Seizure Onset Zone: Applications within Intraoperative and Postoperative Theaters

For drug-resistant epilepsy, surgical intervention, the removal of the brain region that contributes to the seizure onset zone (SOZ), is considered the best action. Intracranial EEG (iEEG) monitoring is used to locate the SOZ by analyzing multiple seizures and their patterns over days of monitoring. However, this is an exhaustive labor-intensive process associated with various risks. High-frequency oscillations (HFOs) of iEEG between 80-600 Hz have emerged as a potential biomarker for SOZ localization. Yet, the presence of pseudo-HFOs, arising from transient artifacts, limited the accurate detection and localization of SOZ using HFOs. This thesis aims to overcome the limitations in distinguishing pathological HFOs from pseudo-HFOs by developing computational tools using sparse signal processing and machine learning techniques. We hypothesized that, unlike pseudo-HFOs, real-HFOs possess a distinct signal characteristic that can be effectively represented using a small number of predefined oscillatory atoms. When applied to iEEG data (n=16) recorded in the epilepsy monitoring unit (EMU), sparse representation-based classification (SRC) framework using a redundant Gabor dictionary achieved 96.65% HFO classification accuracy whereas a previously established baseline method employing time-frequency analysis yielded 82.3% accuracy. Achieved discrimination capacity resulted in a 17.6% improvement in SOZ localization. Here, we approximated the candidate HFO events as a whole block using an analytical Gabor dictionary including atoms at various frequency scales. We further improved this approach by representing the candidate events locally at various temporal scales using an adaptive sparse local representation (ASLR) strategy. Moreover, instead of relying on a predefined analytical Gabor dictionary, we constructed a redundant dictionary automatically from the data using a cascaded residual-based dictionary learning (CRDL) framework. We tested this approach on the iEEG recorded intraoperatively (n=31) that was heavily corrupted by the movements of neurosurgeons and clinical personnel and artifacts originating from nearby devices. We achieved 91.3% classification accuracy in distinguishing between real and pseudo-HFOs while the previously established method had 89.9% accuracy. More importantly, our classification system improved the SOZ localization accuracy up to 70% in certain cases with an average improvement of 19%. As a final step, we converted the whole offline analysis framework into a real-time process, enabling simultaneous recording and HFO analysis. Since the whole framework can be executed online, it can serve as a useful clinical tool for the early detection of the SOZ during or immediately after the implantation of the electrodes. In conclusion, the thesis contributes towards improving real-HFO detection and, as a result, localizing the SOZ effectively. Furthermore, the results emphasize the potential of utilizing sparse signal processing approaches as a valuable tool in the early localization of the SOZ, ultimately aiding in the development of more effective treatment strategies for individuals suffering from drug-resistant epilepsy.