Deep Learning-Predicted RNFL Loss and Incident Glaucoma in the Canadian Longitudinal Study on Aging.

OBJECTIVE

To characterize longitudinal retinal nerve fiber layer (RNFL) thinning predicted from fundus photographs by a machine-to-machine (M2M) model in the Canadian Longitudinal Study on Aging and examine its association with incident glaucoma.

DESIGN

Prospective, population-based cohort study of 18,247 participants (30,202 eyes) aged 45-86 years from 11 Canadian sites. Baseline data from 2012-2015, follow-up through 2015-2018.

METHODS

Fundus photographs analyzed with an optical coherence tomography (OCT)-trained M2M algorithm at baseline and after a 3-year follow-up to estimate RNFL thickness change. Demographic factors, intraocular pressure (IOP), and corneal hysteresis (CH) were assessed.

MAIN OUTCOME MEASURES

Annual rate of predicted RNFL thickness change and risk of incident glaucoma. Linear mixed-effects models identified predictors of RNFL thinning. Incident glaucoma was defined as new self-reported diagnosis at follow-up. Risk factors were evaluated using cox proportional hazards models.

RESULTS

Predicted RNFL loss was faster in glaucomatous versus non-glaucomatous eyes (-0.46 ± 2.28 vs -0.18 ± 2.07 μm/year; P < 0.001). In multivariable analysis, faster thinning was associated with older age (β=-0.223 μm/year per decade; P < 0.001), higher baseline IOP (β=-0.020 μm/year per mmHg; P < 0.001), lower CH (β=-0.026 μm/year per mmHg lower; P < 0.001), and thicker baseline predicted-RNFL (β=-0.665 μm/year per 10 μm; P < 0.001). A significant age×IOP interaction (β=-0.007 μm/year per mmHg per decade; P = 0.036) indicated greater IOP-related thinning in older participants. Among 17,552 participants without baseline disease, 344 (1.9%) converted to glaucoma. Faster predicted-RNFL loss was independently associated with incident glaucoma during follow-up (HR = 1.125 per 1 μm/year increase, 95% CI 1.070-1.183, P < 0.001).

CONCLUSIONS

Deep learning-derived RNFL estimates from fundus photographs were associated with clinically meaningful structural change and predicted incident glaucoma in a population-based cohort. These findings highlight the potential of fundus-based deep learning models to enable glaucoma risk stratification in settings where OCT is impractical or unavailable.