Abstract
Purpose :
Retinal implants aim to restore basic functionality in photoreceptor-degenerated retinas. Towards a quantitative evaluation of the vision restoration we asked how well the stimulated ganglion cell spiking can be predicted. Here we adopted the well-known linear-nonlinear-poisson model, which has been shown to predict the light-stimulated ganglion cell spiking in healthy retina to a relatively high degree.
Methods :
We stimulated ex vivo mouse retina (C57Bl6/J and rd10) using a time-continuous electrical stimulus (low-pass filtered Gaussian white noise) while simultaneously recording the ganglion cell spiking. Electrical stimuli were delivered using an arbitrary subset out of 1024 electrodes embedded in a CMOS-based electrode array. The ganglion cell spiking was identified using 4225 densely packed recording sites, which also allowed to map the axons of activated ganglion cells. The artefact caused by electrical stimulation could be filtered out revealing the stimulation-induced spikes.
Results :
The mean stimulation intensity (linear filter stage) before each spike, the spike-triggered average, showed a distinct peak at 2-5 milliseconds preceding the spike, both in wild-type and in photoreceptor-degenerated rd10 retina. The linear filter together with a static non-linearity was used to predict the stimulated ganglion cell response with high accuracy (~ 60 % of variance explained). The prediction quality decreases with distance of the ganglion cell from the stimulation electrodes. Ganglion cells outside the stimulation areas but with axons crossing these areas were not stimulated.
Conclusions :
Here we demonstrate the applicability of the LNP model for retinal prosthetics being able to predict stimulated spiking activity. We expect this framework to guide the search for efficient and reliable stimuli.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.