AI-based prediction of cone- and rod-function based on retinal microstructure in geographic atrophy secondary to age-related macular degeneration

Abstract

Purpose : To investigate the association of retinal microstructure with cone- and rod-function in geographic atrophy (GA) secondary to age-related macular degeneration (AMD) using artificial intelligence (AI) algorithms.

Methods : Mesopic, dark-adapted (DA) cyan and red sensitivity were measured using fundus-controlled perimetry (“microperimetry”). Test-points were placed along "iso-hulls" at pre-defined distances to the GA boundary using patient-tailored perimetry grids. Retinal microstructure was assessed through spectral-domain optical coherence tomography (SD-OCT) imaging. Fundus-controlled perimetry data were registered to SD-OCT data based on vessel bifurcations. Reflectivity and thickness values were extracted for six retinal layers for each test-point. Using random forest regression, we evaluated (i) the cross-validated (CV) mean absolute error (MAE) with and without patient-specific training data and (ii) increase in out-of-bag mean-squared error (%IncMSE) as measure of SD-OCT feature importance.

Results : Thirty eyes of 30 patients (76.4±7.1 years; 16 female) with GA from the prospective natural progression study DSGA (Directional Spread in Geographic Atrophy; NCT02051998) and 40 normal eyes form 40 age-similar subjects were included. For patients with GA, sensitivity was predicted with a MAE [95% CI] of 4.15 dB [3.39; 4.91] for mesopic, 5.46 dB [4.64; 6.27] for DA cyan and 3.98 dB [3.44; 4.53] for DA red testing in absence of patient-specific data. Partial addition of patient-specific sensitivity data to the training sets decreased the MAE to 2.53 dB [2.49; 2.58], 3.21 dB [3.16; 3.26] and 2.55 dB [2.52; 2.58]. For all three types of testing, the outer nuclear layer thickness constituted the most important feature (70.88, 78.66 and 107.81 %IncMSE). AI-based spatial mapping of “inferred sensitivity” across the whole retina imaged by SD-OCT and comparison to normal data revealed that DA cyan sensitivity loss spatially exceeded mesopic sensitivity loss in eyes with GA.

Conclusions : AI-based “inferred sensitivity” mapping may be applicable as quasi-functional clinical outcome to (partially) substitute for time-consuming psychophysical testing. Rod-function appears to be more severely affected in eyes with GA secondary to AMD and may constitute a potential therapeutic target to prevent downstream cone-degeneration.

This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.