Purchase this article with an account.
Moosa Zaidi, Gorish Aggarwal, Nishal P. Shah, Orren Karniol-Tambour, Georges Goetz, Sasi Madugula, Alex R. Gogliettino, Eric G. Wu, Alexandra Kling, Nora Brackbill, Alexander Sher, Alan M. Litke, E.J. Chichilnisky; Inferring retinal ganglion cell light response properties from intrinsic electrical feature. Invest. Ophthalmol. Vis. Sci. 2021;62(8):3167.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
An ideal epiretinal implant for treating vision loss would stimulate each retinal ganglion cell (RGC) of each distinct type according to its natural function, to reproduce the neural code of the retina. However, no methods exist to determine the normal, healthy light response properties of each RGC of each type in a region of retina that is no longer light-responsive. This study presents a method to infer a model of RGC light responses using only intrinsic features of recorded electrical activity.
Analysis was performed on 283 recordings from populations of ON and OFF parasol RGCs in isolated macaque retina using 512-electrode arrays. The type of each recorded RGC was determined from its light responses. To emulate interfacing to a blind retina, a classifier was used to infer the type (ON or OFF) of each cell using two features of recorded electrical activity, without reference to the light stimulus: the average electrical image (EI) of a spike over the electrode array, and the autocorrelation function (ACF) of spike times. EIs were also used to infer receptive field locations. A linear-nonlinear Poisson (LNP) light response model was then inferred for each cell using its true cell label, its estimated location, and the average spatiotemporal filter and nonlinearity in the population. The prediction accuracy of inferred models and models fitted to measured responses was compared, for white noise and natural scenes stimuli (Fig. 1).
K-means (k=2) clustering of the top 2 principal components of ACFs reliably separated ON and OFF parasol cells in individual retinas. However, cluster labels (ON, OFF) could not be reliably determined from ACFs, due to inter-retina variability. Instead, ACF clusters were labeled using “votes” from a cell-by-cell linear classifier of selected EI features. This approach achieved 94% mean accuracy on 25 test retinas. Across six retinas, the RGC light response models produced a mean correlation between inferred models and data of 0.48 and 0.51, and between fitted models and data of 0.63 and 0.57 (an upper bound), for white noise and natural scenes stimuli respectively.
Intrinsic features of electrical activity can be used to infer RGC light response models.
This is a 2021 ARVO Annual Meeting abstract.
Figure 1. Example rasters of light responses measured (data), predicted from an inferred light response model (infer), or predicted from a light response model fitted to recorded data (fit).
This PDF is available to Subscribers Only