Fundus Refraction Offset as a Personalized Biomarker for 12-Year Risk of Retinal Detachment

Abstract

Purpose: The purpose of this study was to investigate the potential of a novel anatomical metric of ametropia-fundus refraction offset (FRO)-in stratifying the risk of retinal detachment (RD) or breaks, beyond the influence of risk factors including spherical equivalent refraction (SER).

Methods: Participants from the UK Biobank with no prior history of RD/breaks were analyzed (n = 9320). The onset of RD/breaks over a 12-year follow-up period was determined based on linked healthcare data. A previously trained deep learning model was applied to each fundus photograph to predict SER. FRO was defined as the error in the fundus-predicted SER, with a negative value indicating a relatively myopic-looking fundus. Cox regression was used to examine the association of baseline FRO with RD/breaks-adjusting for baseline SER, baseline age, sex, and cataract surgery during follow-up. In a subgroup of participants (n = 7127) with high-quality optical coherence tomography scans, we additionally adjusted for baseline macular thickness (MT). All analyses initially considered any RD/breaks as the event, followed by rhegmatogenous RD/breaks.

Results: The mean (SD) baseline age was 54.8 (8.2) years. Sixty-four participants developed RD/breaks (of any subcategory), with a mean (SD) of 7.0 (3.3) years between baseline and disease onset. A more negative baseline FRO was independently associated with an increased risk of any RD/breaks (adjusted hazard ratio [HR] = 0.66, 95% confidence interval [CI] = 0.50-0.87, P = 0.003) and rhegmatogenous RD/breaks (HR = 0.61, 95% CI = 0.45-0.82, P = 0.001). Similar independent associations were evident in the subgroup analysis that additionally adjusted for MT.

Conclusions: A more negative baseline FRO is associated with a higher risk of developing RD/breaks, even among individuals with similar baseline SER and other risk factors. This demonstrates a potential benefit of shifting towards an anatomic definition of myopia.

Figure 1.

Overview of fundus refraction offset (FRO). A deep learning model was previously trained on fundus images from phakic eyes without any posterior eye conditions to predict spherical equivalent refraction (SER), with the expectation that the trained model—having achieved an optimal bias-variance trade-off (good generalizability to unseen images)—would have learned to capture the non-pathological variations in fundus appearance across a broad spectrum of SER. When the trained model is applied to unseen images, the fundus-predicted SER is termed fundus equivalent refraction (FER) to distinguish it from SER in the strictly one-dimensional and on-axis sense. As FER is derived directly from pixel-level information across an individual's fundus, it is thought to contain “higher-dimensional” information about the eye, capturing more subtle interindividual differences in posterior segment anatomy not otherwise fully captured by SER, as demonstrated in our recent work. To derive FRO, the deviation of each observation from the line of best fit (i.e. residuals from linear regression) is computed, as visually represented by the vertical dashed lines in the scatterplot. This is illustrated using three examples with the same SER but different FER: the eye that lies on the line (datapoint B) shows zero deviation from the norm; the eye that shows positive deviation (datapoint A) has a relatively hyperopic fundus appearance; and the eye that shows negative deviation (datapoint C) has a relatively myopic fundus appearance.

Figure 2.

Kaplan-Meier survival curves showing the unadjusted survival probability, that is, the probability of remaining event-free over time, in participants with different baseline categories of FRO. Event is defined as any subcategory of retinal detachment or retinal breaks on the left, whereas on the right only rhegmatogenous retinal detachment or retinal breaks. “Number at risk” indicates the number of participants that remain event-free or have not been right-censored at each time point.