Abstract
Purpose :
There is a growing body of research linking vasculature-related retinal traits computed from fundus images (e.g. Fractal Dimension, tortuosity) and health-related attributes beyond the eye (e.g. brain imaging phenotypes, cardio- and neurovascular disease outcomes). These studies typically exclude many images (in the order of 20-40%) that are of insufficient quality to compute retinal traits. We set out to investigate whether image quality exclusions might induce a selection bias.
Methods :
We used logistic regression to examine the relationship between key attributes of 82,274 UK Biobank participants (age, sex, BMI, blood pressure, smoking status (ever vs never), ethnicity (White vs non-White), social deprivation) and image quality exclusion as a binary indicator that no fundus image of sufficient quality was available. We used a quality score introduced by recent work on Fractal Dimension and cardiovascular disease. It was designed to decide whether an image can be analysed by the VAMPIRE software.
Results :
Summary statistics and the results of the logistic regression are reported in Tables 1 and 2, respectively. Being male, non-white, older, higher BMI and higher blood pressure were all associated with a statistically significant (p<0.05) increase in the risk of being excluded. Effect sizes were particularly large for age (Odds Ratio [95% CI]: 1.0607 [1.0587-1.0629], i.e. an additional decade increases odds by 80% [77-84]), non-White ethnicity (OR: 1.1522 [1.0986-1.2092]), and male sex (OR: 1.1444 [1.1096-1.1794]).
Conclusions :
Our results suggest that image quality exclusions induce a selection bias. As a result, health research using retinal traits might not generalise to the population of UK Biobank, let alone the whole UK. Concerningly, the exclusions exacerbate the already poor representation of non-White ethnicities in health research when controlling for other variables.
This has two key implications: First, researchers using retinal traits should be aware that image quality exclusions likely induce a selection bias and report how population statistics change as a result. Second, more robust methods for computing retinal traits should be developed so that fewer images need to be excluded in the first place.
Future work should quantify the impact of these exclusions on the down-stream analyses and investigate the causal mechanisms of the observed relationships.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.