Abstracts

Rationale: Introduction: Epilepsy is a common and debilitating neurologic disease affecting 1% of the population characterized by recurrent seizures. In the United States, approximately 35% of those afflicted cannot achieve reasonable control of their seizures on anti-epileptic medications. The device world has taken notice and in 2014 published results from the Pivotal Trial of a responsive neural-stimulation clinical series. Missing from these devices is information about seizure onset channels or any localization at all. All recording channels are weighted equally in the closed-loop warning system. With this in mind, an automated method for seizure onset detection to be implemented into such devices would be very useful for the epilepsy and medical device communities. Methods: Materials and Methods: In a series of patients with temporal lobe focal epilepsy at Mayo Clinic, analysis on 2-hour segments of interictal (between seizure) data was taken in n=7 patients. Data was sampled at 32,000 samples/sec and decimated to 5,000. Anti-aliasing filtering was done at 1,000 Hz. Primary spectral analysis was performed between standard clinical Berger Bands (0-3 Hz = Delta, 3-8 Hz = Theta, 8-13 Hz = Alpha, 13-25Hz = Beta) as well as higher oscillations frequencies [Staba 2004] 25-55Hz = Low Gamma, 65-100Hz = High Gamma, 100-150Hz = epsilon gamma, >150Hz taken as ripple frequency range. After pre-processing using ICA subtraction methods, and accounting for any events (seizures were no closer than 12 hours from any 2-hour segment analyzed) or discontinuities, analysis was performed in both asleep and awake states. Feature extraction was undertaken using methods previously employed [Varatharajah 2016] for the following features: Power-in-Band (PIB) utilizing log-scaled weightings of bands to account for the power law, phase-amplitude coupling [Berry 2016], spectral coherence (SPCO), and time correlation (TMCO). Using this combination of features, coalescent analysis was undertaken on PAC and sequential analysis was undertaken based on a clustering approach. Additionally, a sequential analysis was undertaken based on a Bayesian approach. Gold standard SOZ channels were determined by a trained epileptologist and ROC was utilized to assess the validity of the approach for each patient at different thresholds. Results: Results and Discussion: PAC high value epoch analysis resulted in a AUC of 0.79 (range 0.5-0.9). Clustering based approaches resulted in AUC in this series of patients 0.77 with range 0.64-0.92. There were certain patients in this series in whom the approach was highly sensitive and specific and others whose seizure onset localization was barely better than chance. This suggests that future work will need to be tailored to the specific features in that patient which result in optimal localization. Conclusions: Conclusions: This work shows that automated localization is possible and may prove useful in devices which track not just electrophysiologic biomarkers of seizure onset but also any neurologic disorder which can be tracked with electrophysiology. Funding: Acknowledgements:This project was supported by the NIH NINDS (U01-NS073557, R01-NS92882), and the Mayo Graduate School SURF Program (YV). References: Staba RJ Ann Neurology 2004 Jul;56(1):108-15. High-frequency oscillations recorded in human medial temporal lobe during sleep. Varatharajah Y IJNS 2016 Jul. Pre-ictal time, postictal effects, and seizure clustering in Ieeg-based seizure prediction in naturally occurring epilepsy. Berry BM in press. Phase-Amplitude Coupling as an interictal signature of seizure onset zone.