Abstract
Purpose :
By electrical stimulation on retina, there are two - early (directly-evoked) and late (indirectly-evoked) - retinal ganglion cell (RGC) responses evoked. Direct RGC response which is desirable for the retinal prosthesis to recreate the appropriate spike pattern is likely to be obscured by stimulus artifact. In this study, we aim to isolate the directly-evoked RGC response from stimulus artifact using three algorithms; Subtraction of Artifacts by Local Polynomial Approximation (SALPA), Moving Average Filter (MAF), and Forward Reverse (FR) filter. SALPA is a well-established algorithm and the other two are new custom-made algorithms for artifact subtraction.
Methods :
Rd1 mice after postnatal 10 week were used (n=4). The RGC responses were extracellularly recorded with 8 × 8 multi-electrode array (MEA) in which one electrode was used as stimulating electrode and all other electrodes as recording electrode. Fifty cathodic phase-1st biphasic current pulses (duration 500 μs, amplitude 5~60 μA) were applied at every 1 sec. Three artifact subtraction algorithms were used: Subtraction of Artifacts by Local Polynomial Approximation (SALPA) is based on curve fitting of polynomial function. For Moving Average Filter (MAF), fitting number of 23 was used. For forward reverse (FR) filter, 100 Hz high pass and 500 Hz low pass was used for cut-off frequency. The latencies of first spikes following 50 stimuli were averaged in each electrode and the latencies acquired from three algorithms were statistically compared. False positive error (algorithm mistakes the artifact as the spike) and missing error (missing early response as residual artifact) were calculated.
Results :
SALPA, MAF, and FR filter detected latency of early RGC spikes as 8 ± 3.6 ms, 13.1 ± 5.2 ms and 13.2 ± 4.7 ms, respectively (p < 0.001). False positive error and missing error for SALPA, MAF, and FR filter were 4%, 5%, and 2%, and 2%, 26%, and 21%, respectively.
Conclusions :
In terms of latency of early RGC response, SALPA detects latency earlier than two other algorithms (p < 0.001). For false-positive error, FR filter shows best performance. For missing error, SALPA shows best performance. Since MAF and FR filter’s performance depends on fitting number and cut-off frequency, in future study, we try to minimize the false-positive and missing error by fine-tuning of both MAF and FR filter.
This is an abstract that was submitted for the 2016 ARVO Annual Meeting, held in Seattle, Wash., May 1-5, 2016.