Abstracts

Rationale: While electrical stimulation is a promising therapy for refractory epilepsy, selection of stimulation settings poses a significant challenge and may lead to inconsistent and inconclusive results. As a case in point, despite strong suppression via optogenetic manipulation, electrical stimulation of the cerebellum for seizure control has had inconsistent results. We hypothesize that effective seizure-control via cerebellar electrical stimulation is dependent on stimulation settings. To identify effective parameters from hundreds of possible combinations in a data-driven manner, we incorporated Bayesian optimization, a rational and transparent search algorithm, into on-demand seizure-intervention (Fig 1).

Methods: Seizures in the intrahippocampal kainate mouse model of temporal lobe epilepsy were detected online from local field recordings from the hippocampus, and on-demand electrical stimulation was delivered to the midline cerebellum (vermis), either parallel or perpendicular to the longitudinal axis. Using Gaussian process regression and Bayesian optimization, we varied charge amplitude (-75 to 75 nC), frequency (2-512 Hz), and pulse width (100-500 us), using two different sets of step sizes (and associated length constants), resulting in a total of either 1,081 or 271 unique possible combinations. Through optimization, a parameter set that was associated with the shortest seizure durations was identified for each mouse, and then tested head-to-head against no intervention.

Results: Bayesian optimization from 1,081 possible combinations successfully identified settings that significantly inhibited seizures, for both the perpendicular (n=6/6 p< 0.05; 68±9.5% reduction in average seizure duration) and parallel (n=6/6 p< 0.05; 56±13.8%) electrode orientation (Fig 2). Optimization within the second parameter space (271 combinations) also identified effective parameter settings (perpendicular: n=6/6 p< 0.05, 77±4.9% reduction; parallel: n=3/4 p< 0.05, 40±8.1% reduction). In contrast, there was never seizure inhibition when an identified “not-optimal” parameter set was used. The identified optimal combination was unique for each mouse, but followed similar trends of higher frequency and charge amplitude. We therefore tested settings based on group data, and found that group average optimized settings were as effective in reducing seizure duration as individual optimized settings (perpendicular: n=6; parallel: n=3).