AI-derived OCT features track inter-eye contrast-sensitivity differences

Abstract Purpose AI-based OCT analysis can predict visual function across individuals, but these associations may reflect shared participant characteristics rather than eye-specific information. Comparing fellow eyes tests whether OCT features capture functional differences within the same person. The purpose is to test whether inter-eye OCT feature differences track contrast-sensitivity (CS) differences after visual-acuity (VA) adjustment. Methods We analyzed 2,015 AI-READI participants with bilateral OCT and visual-function measurements. A fixed RETFound encoder extracted OCT features. Photopic and mesopic ridge models were developed in training (n = 1,412) and applied unchanged to validation (n = 307) and test (n = 296). Analyses included eye-swap and outcome-permutation controls, VA adjustment, and exploratory retinal-thickness comparisons. Results Photopic correlations between predicted and observed inter-eye CS differences were 0.334 in validation and 0.318 in test; predictive R² values were 0.095 and 0.101. Mesopic correlations were 0.373 and 0.194; predictive R² values were 0.133 and 0.027. All associations exceeded control null expectations. VA-adjusted partial correlations were 0.272 (95% CI, 0.174–0.364) for photopic and 0.171 (95% CI, 0.069–0.264) for mesopic CS. AI features yielded numerically higher photopic correlations than thickness features; combining features did not consistently improve performance. Conclusion AI-derived OCT features tracked inter-eye CS differences after VA adjustment, with more consistent photopic performance. Translational Relevance: AI-derived OCT features may complement visual acuity in assessing contrast sensitivity.