Artificial intelligence for morphology-based function prediction in neovascular age-related macular degeneration

Abstract

Spatially-resolved mapping of rod- and cone-function may facilitate monitoring of macular diseases and serve as a functional outcome parameter. However, mesopic and dark-adapted two-color fundus-controlled perimetry (FCP, also called “microperimetry”) constitute laborious examinations. We have devised a machine-learning-based approach to predict mesopic and dark-adapted (DA) retinal sensitivity in eyes with neovascular age-related macular degeneration (nAMD). Extensive psychophysical testing and volumetric multimodal retinal imaging data were acquired including mesopic, DA red and DA cyan FCP, spectral-domain optical coherence tomography and confocal scanning laser ophthalmoscopy infrared reflectance and fundus autofluorescence imaging. With patient-wise leave-one-out cross-validation, we have been able to achieve prediction accuracies of (mean absolute error, MAE [95% CI]) 3.94 dB [3.38, 4.5] for mesopic, 4.93 dB [4.59, 5.27] for DA cyan and 4.02 dB [3.63, 4.42] for DA red testing. Partial addition of patient-specific sensitivity data decreased the cross-validated MAE to 2.8 dB [2.51, 3.09], 3.71 dB [3.46, 3.96], and 2.85 dB [2.62, 3.08]. The most important predictive feature was outer nuclear layer thickness. This artificial intelligence-based analysis strategy, termed “inferred sensitivity”, herein, enables to estimate differential effects of retinal structural abnormalities on cone- and rod-function in nAMD, and may be used as quasi-functional surrogate endpoint in future clinical trials.