Abstracts

Rationale: There are currently two FDA-approved implantable devices for electrical stimulation of the brain (ESB) for epilepsy, but only approximately 15% of patients using these devices report extended seizure free periods ( > 6 months). Pivotal trial seizure outcomes were based on patient diaries, which are notoriously inaccurate. The development of ESB devices with both sensing and stimulation capabilities opens the possibility of accurate, automated seizure diaries to guide therapy optimization. Here we report initial efforts to develop automated electronic seizure catalogs. In this study, we use a power-in-band (PIB) thresholding technique that can be embedded on devices to distinguish between seizure and non-seizure states. This simple algorithm can be implemented on the investigational Medtronic Summit RC+S^TM.

Methods: We implanted four patients with the investigational Medtronic Summit RC+S^TM sense and stimulation device with leads in bilateral hippocampi (Hc) and anterior thalamic nuclei (ANT) and have investigated automated ANT seizure detection in two patients. We collected long-term, ambulatory intracranial EEG (iEEG) data, scored by board certified epileptologists, and implemented an in silico, Python mockup of a PIB threshold classifier embedded on the Summit RC+S^TM. We systematically tested the performance of this ANT seizure classifier using PIB features with central frequencies from 3 to 98 Hz (5 Hz bandwidth) and time windows from 2 seconds to 10 minutes. Area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (AUPRC) metrics were used to compare performance between parameters. Parameters were optimized on the first patient and tested on the second.

Results: We recorded 46.3 days of stimulation-free iEEG from the two subjects for analysis, including 20 and 3 ANT recorded seizures from each patient, respectively. We determined that a 7 Hz center frequency in 10-second windows was optimal for seizure detection and was able to obtain AUROC=0.997 and AUPRC=0.59 with these parameters for the first patient, with the 80% sensitivity achieved at the expense of 1 false positive per day. For the out-of-sample patient, 100% sensitivity was achieved with 0 false positives.

Conclusions: Our results show that a detection algorithm that should be compact enough to be embedded in commercial implantable devices is promising. While these parameters were optimal for the two patients analyzed, the generalization of this algorithm remains unclear, especially in cases during stimulation. Significantly, implementation of an embedded seizure detection algorithm on implantable devices will be valuable closed-loop neurostimulation and optimizing epilepsy management.

Funding: Please list any funding that was received in support of this abstract.: NIH UH2/UH3 NS095495: Neurophysiologically Based Brain State Tracking & Modulation in Focal Epilepsy.