Predicting Retinal Nerve Fiber Layer Thickness From Ocular Hypertension Treatment Study Optic Disc Photographs

Importance: Deep learning predictions of retinal nerve fiber layer (RNFL) thickness derived from optic disc photographs may help to determine risk for development of primary open-angle glaucoma (POAG) in patients with ocular hypertension. Objective: To predict mean RNFL thickness from the optic disc photographs from the Ocular Hypertension Treatment Study (OHTS) and assess the utility of predicted RNFL thickness as a risk factor for the development of POAG. Design, Setting, and Participants: This diagnostic study evaluated 3272 eyes from 1636 participants with ocular hypertension but without POAG at the time of enrollment in the OHTS 1 and 2 trials. The OHTS was a multicenter study, with OHTS 1 and OHTS 2 collectively extending from February 28, 1994, to December 30, 2008. Optic disc photographs, baseline demographics, and clinical examination findings were included in the analysis. An OCT-trained deep learning model (machine-to-machine [M2M] model) was used to generate predicted RNFL thicknesses from 66 714 optic disc photographs. Main Outcomes and Measures: The primary outcomes were factors (including predicted RNFL) that correlated with conversion to POAG from the OHTS cohort, identified by proportional hazards models. Results: Among 1444 participants with ocular hypertension from the OHTS cohort, mean (SD) age was 56.0 (9.5) years, and 833 participants (57.7%) were female. Mean (SD) baseline predicted RNFL was 94.1 (7.1) μm for eyes that converted to POAG and 97.1 (7.0) μm for eyes that did not convert to POAG (mean difference, 3.0; 95% CI, 2.2-3.8; P < .001). Predicted baseline RNFL was a predictor of conversion to POAG during follow-up in Cox proportional hazards models in univariable analysis (hazard ratio [HR], 1.97; 95% CI, 1.60-2.42; P < .001) and multivariable analysis (HR, 1.83; 95% CI, 1.49-2.25; P < .001) per 10-μm thinner in predicted RNFL. Baseline age, intraocular pressure, central corneal thickness, pattern standard deviation, mean deviation, and cup-disc ratio remained predictors of conversion to POAG in both univariable and multivariable analysis. Longitudinal change in predicted RNFL (per 1-μm/year faster loss) was also a predictor of conversion to POAG (HR, 6.01; 95% CI, 3.33-10.64; P < .001). Conclusions and Relevance: In this diagnostic study, baseline M2M-predicted RNFL thickness and longitudinal rate of change in predicted RNFL were putative risk factors for the development of glaucoma in patients with ocular hypertension. These findings support the utility of M2M-predicted RNFL thickness to assess baseline glaucoma risk and monitor for glaucoma progression.