Abstracts

Rationale: Electrical brain stimulation (EBS) is an established therapy for drug-resistant epilepsy. Novel implantable neural sense and stimulation devices (INSS) enabling continuous intracranial electroencephalographic (iEEG) streaming provide new opportunities for objective outcome evaluation with seizure diaries and sleep scoring. However, new challenges arise when collecting iEEG data with concurrent EBS. EBS-induced stimulation artifacts and brain electrophysiology changes are factors affecting the performance of automated classification algorithms. We investigated the feasibility of utilizing iEEG data to build an automated behavioral state classifier under different EBS paradigms.

Methods: Four human subjects underwent chronic ambulatory monitoring using the Medtronic investigational Summit RC+S^TM INSS device with electrodes implanted in bilateral hippocampus (HC) & anterior nucleus of the thalamus (ANT). Standard sleep clinical polysomnography (PSG) and continuous iEEG data streaming from the INSS were recorded simultaneously over three nights in the epilepsy monitoring unit. Different EBS parameters (2, 7 and 145 Hz) were trialed in 5–15-minute-long epochs with wash-out no-stim periods. PSG data were scored according to gold standard sleep categories using AASM2012 rules. An automated classification algorithm was designed and trained using only no-stim iEEG data and tested under the 2, 7, and 145 Hz EBS setups.

Results: An automated behavioral sleep state iEEG classifier (wake, rapid eye movement (REM and non-REM) had the overall average F1-score was 0.889 across all modes of stimulation. The models were deployed on long-term data (over 30 months of continuous iEEG in total) to create a sleep/wake profile for all patients.

Conclusions: The trained sleep classification models enable the assessment of behavioral states under different low (2&7 Hz) and high ( >100Hz) frequency EBS for four human subjects implanted with Medtronic investigational Summit RC+S^TM INSS device for epilepsy treatment. The study shows that the behavioral state classification models can be trained using no-stim data, resulting in high classification rates in EBS.

Funding: Please list any funding that was received in support of this abstract.: This research was supported by National Institutes of Health (UH2&3-NS95495), DARPA Morepheus, LQ1605 from the National Program of Sustainability II (MEYS CR, Czech Republic), and institutional resources from Mayo Clinic, Rochester MN USA, and Medtronic Plc, Minneapolis, MN, USA.