In 2011, 2.71 million people carried a diagnosis of Primary Open Angle Glaucoma (POAG) in the United States alone.¹ The percentage of individuals with POAG was significantly higher amongst individuals between 70 and 79 years, females, and non-Hispanic Whites.¹ By the year 2050, the number of individuals carrying a diagnosis of POAG is projected to rise to 7.32 million — indicating a rise in prevalence of 28% every decade.¹ The largest increases are expected among Hispanic Americans, indicating a shift towards an increased burden of glaucoma among minority communities. This trend is not new as several studies have historically shown that racial and ethnic minority communities have the highest rates of glaucoma when compared to White communities. The Baltimore Eye Survey was the first to reveal that Black Americans had an age-adjusted glaucoma prevalence 4–5 times higher than White Americans.² A study in Salisbury, Maryland found White participants over age 75 had a glaucoma prevalence of 9.4% while Black participants in the same age group had a much higher glaucoma prevalence rate of 23.2%.³ Moreover, after adjusting for race and medical co-morbidities, people with lower income had a higher incidence of glaucoma compared to those with higher income.⁴ There is a compelling need to better understand the social, cultural, institutional, and environmental factors that lead to such stark inequities in glaucoma diagnosis in the United States.

As glaucoma can be asymptomatic in its earliest stages, without screening, glaucoma often remains undetected until late stages when central vision is affected, and people seek care. It is estimated that 10–33% of individuals with glaucoma already have significant visual field loss upon first presentation to an ophthalmologist.⁵ One approach to mitigate disparities in the diagnosis of glaucoma involves early identification of the disease through screening programs targeted at communities at high-risk of disease, such as communities with large proportions of people living in poverty, or large proportions of minorities. Identifying disease earlier opens the pathway towards earlier initiation of treatment. Optical Coherence Tomography (OCT) offers substantial potential as a screening tool for such community screening initiatives. OCT is a noninvasive, high-resolution imaging technique that enables real-time measurement of tissue thickness in the posterior pole. Unlike other diagnostic tools, such as visual fields or fundus photography, OCT is fast to administer, is less demanding of patients, and has good reproducibility.^6–9 Automated assessment of retinal thickness not only provides a reliable method to track disease progression, but also facilitates a comparison with healthy individuals for disease identification. Optic nerve head (ONH) parameters and retinal thicknesses are contrasted with healthy normative values in the OCT device’s proprietary software to generate a color-coded label for identification of “normal” versus “abnormal values”. This color-coding system provides a simple heuristic aid for glaucoma decision-making, promoting more streamlined and consistent medical care.¹⁰

Before OCT can be implemented into large community-based screening protocols, it is critical to assess the existing reference databases that define the distinction between “normal” and “abnormal” eyes. These databases need to be appropriate for a diverse range of populations. There are many manufacturers of OCT machines, and each configures their software with their own proprietary normative dataset. For instance, the reference databases for the Spectralis (Heidelberg Engineering, Inc., Germany) and Cirrus (Carl Zeiss Meditec, Inc., Dublin, CA) machines, two of the most commonly used machines in the United States, include subjects with no known eye disease and specifically exclude those with visual field damage, high intraocular pressure (IOP), and severe myopia (−7 diopters or more for Spectralis, and −10 diopters for Cirrus).¹¹ Further, the Spectralis normative database predominantly includes White individuals (79%), and White individuals make up 43% of the Cirrus normative database. The Spectralis cohort excludes those with a family history of glaucoma, and Cirrus excludes those with systemic hypertension and diabetes. Most notably, only 284 and 330 eyes are included in the Cirrus and Spectralis databases, respectively.¹¹ Thus, while the number and demographic composition of the individuals in these reference databases vary, OCT normative cohorts typically consist of small, racially and ethnically homogenous samples, that are often without systemic conditions.^11,12

To better understand the applicability of OCT technology to distinguish healthy individuals from those with glaucomatous disease in real-world community contexts, this study aims to investigate differences in ONH parameters and retinal thicknesses between healthy individuals from the Topcon 3D OCT-1 Maestro normative database and healthy individuals from a population with a higher proportion of people identifying as Black or African American and a higher proportion of people with low socio-economic status. These individuals were part of a cohort enrolled in a free eye disease screening program in Southeastern Michigan.

The aim of this study was to compare optic nerve head (ONH) parameters and retinal thickness measurements between participants in the Topcon Maestro normative cohort database and a large, diverse group of participants from a community screening program, MI-SIGHT. The average age of the Topcon cohort was 46.3 years, with 22% of the individuals aged between 18–30 years. In comparison, only 9% of our MI-SIGHT dataset was within this age bracket, with an average age of 53.9 years. There were also significant differences in the racial and ethnic makeup of the two cohorts. In the Topcon normative dataset, 59% of the participants identified as White and 20% identified as Black or African American. In the MI-SIGHT dataset, 40% of participants identified as White and 61% as Black or African American. The Topcon normative dataset included participants with manifest spherical equivalents varying from −11.00 to +4.50D, whereas our dataset featured a broader array of refractive error, ranging from −11.5 to +9.5D. The MI-SIGHT healthy cohort did not exclude individuals based on family history of eye disease, current medication use, or systemic conditions, while the Topcon normative cohort excluded those with a history of leukemia, dementia, or multiple sclerosis, or who were using hydroxychloroquine or chloroquine. Lastly, the Topcon normative cohort participants were recruited from academic and optometric centers across the United States, while the MI-SIGHT healthy cohort was recruited from two community clinics who provide care to medically underserved, economically disadvantaged communities in Southeastern Michigan. Considering these differences, we sought to understand how results from the Topcon software could still be applied to the MI-SIGHT healthy older, majority Black/African American cohort.

When comparing OCT parameters between the Topcon normative cohort and the MI-SIGHT healthy cohort, we found almost all ONH parameters did not differ between the two cohorts. However, when stratifying by self-identified race, we found statistically significant differences in all included ONH parameters between MI-SIGHT Black/African American and MI-SIGHT White participants and the predominantly White Topcon normative cohort. Both the White MI-SIGHT participants and the Black/African American MI-SIGHT participants had significantly different vertical cup to disc ratios, cup to disc areas, cup areas, and cup volumes from the Topcon normative cohort. These differences were masked within the overall cohort comparisons, as the Topcon cohort typically fell in between the average values of MI-SIGHT White participants and MI-SIGHT Black/African American participants. For example, while the mean c/d areas were 0.32mm² and 0.24mm² for the Black/African American MI-SIGHT participants and White MI-SIGHT participants respectively, the Topcon normative dataset had a mean c/d area of 0.28mm². These findings are consistent with results from numerous studies utilizing other imaging modalities including Heidelberg Retinal Tomography II, confocal scanning laser ophthalmoscopy, and stereoscopic fundus photographs that have similarly identified larger disc and cup areas in Black individuals compared to White individuals.^16–20 Notably, there appears to be less variation in retinal ganglion cell thickness values between White and Black individuals based on our cohort data and others and RGC thickness perhaps should be further investigated as a diagnostic marker in glaucoma screening.^21–23

Our results also coincide with previously published literature demonstrating significantly thicker RNFL values in the superior and inferior quadrants and thinner temporal quadrants in African American individuals compared to White individuals.^24–29 Accordingly, Black/African American MI-SIGHT participants more frequently exhibited RNFL thicknesses with white and red color labels in the superior and inferior quadrants and less frequently exhibited yellow and green color labels in these quadrants than expected if thickness values were to follow the expected normative distribution, even when compared to White MI-SIGHT participants. Similar results were found by Addis and colleagues when evaluating color code label distributions for 466 healthy eyes from Black/African American participants in the Primary Open-Angle Black American Glaucoma Genetics (POAAGG) study using the Cirrus HD-OCT device.²⁴ They found that more RNFL measurements were assigned as white (9.5%) and red labels (1.7%) than expected and fewer measurements were assigned as green (85.3%) and yellow (3.5%) labels than expected. Unlike the Addis study, we utilized the Topcon Maestro machine, which is more often used in community screening programs than the Cirrus machine, given its more affordable pricing. We also investigated how both individuals self-identifying as Black and/or African American and those self-identifying as White compared to the normative dataset. Systematic over-classification of white and red labels in certain populations is clinically important. Populations with RNFL measurements frequently overclassified with white labels may need to have a larger decrement in RNFL before disease becomes clinically detectable. Conversely, a higher than expected percentage of red labels in certain populations can cause overdiagnosis of disease due to a systemic bias towards classifying these individuals as having unhealthy RNFL.

The discussion around normative databases being race or ethnicity-specific has been ongoing.²¹ Hong et al. evaluated 295 eyes of 166 healthy Korean people and found that the kappa agreement between the original and adjusted color labels for RNFL thickness was between 0.500–0.806. When they applied a Korean-person specific normative database to their cohort, they found that more than 10% of labels were judged differently and the percent of eyes that had yellow or red labels increased from 0.3% to 3.7%, or closer to the expected 5%.³⁰ Perez et al. also found that after applying a White-person specific normative database for the RT-Vue scan, whose normative database is comprised of only 18% White participants, the percent of RNFL abnormal color labels decreased significantly for the overall average thickness in the White participants.²¹ Zelefsky et al. investigated whether a race-specific database would increase the ability of confocal scanning laser ophthalmoscopy to differentiate normal from glaucomatous eyes and found that for White participants, the race specific database increased sensitivity with a comparable level of specificity, but for African-American participants, sensitivity increased but specificity decreased.³¹ In addition, some researchers have looked into adjusting for disc size as a proxy for the variability associated with race. In eyes with larger discs, a constant diameter scan would theoretically measure RNFL locations closer to the disc, and capture thicker RNFL values.³² However, Silverman et al. found that the predominantly White Spectralis reference database performed similar to the more racially and ethnically heterogeneous Cirrus reference database in differentiating normal versus early glaucoma eyes, even without adjustment for disc size.¹² Results from our multivariable multinomial logistic regression similarly shows the variable influence disc area has on predicting RNFL color-code labels.

A novel contribution of this study is the results of our multinomial logistic regression analysis, which demonstrates that RNFL thickness color code classification is influenced by many factors beyond self-identified race or even disc size, two of the most frequently studied predictors of RNFL thickness. Even after controlling for self-identified race and disc area in our analysis, age and SE were predictive of having either a red or white label, compared to green, in the superior quadrant, but neither were predictive in the inferior quadrant. We found that identifying as White race decreased the odds of having a white compared to green color label in the superior quadrant and increased the odds of having a yellow compared to green color label in the inferior quadrant, but otherwise did not significantly affect the probability of having a particular color label. We did not identify an association between having a white color label and having a larger disc in any quadrant except for the temporal quadrant. These results suggest that there is likely a complex interaction between numerous factors, beyond race, that influence RNFL thickness and classification, including refractive error, axial length, corneal astigmatism, systemic conditions such as diabetes and hypertension, retinal vascular density, and optic nerve size and orientation.^24,33 Researchers have even found associations between RNFL thickness measures and unique factors such as having polycystic ovary syndrome, being exposed to ambient air pollution, and mountaineering in high altitude areas.^34–36 The diversity of the human experience extends far beyond the classifications of race and ethnicity and this is reflected in the diverse factors that impact RNFL values.

There are several inherent problems to creating race/ethnicity-specific databases.^37,38 First, there exists significant differences when it comes to broad racial groupings based on American categorizations of “Asian” or “African” or “Other,” as these categories are arbitrary and do not take into consideration country of origin or ancestry and make it difficult when people’s families come from more than one category.^31,39–44 The percentage of people identifying as multi-racial is growing faster than any other racial group per the United States census, increasing from 2.9% in 2010 to 10.2% of the population in 2020.⁴⁵ It hardly seems feasible to provide ethnicity-specific normative databases for every single race or ethnicity or combination thereof. Some have even posited that a correction factor should be included when predicting probability of RNFL thickness labels, as is done with age, to take into account ethnic diversity. Again, this tends to be a narrow view of biologic race as it necessitates categorizing people into pre-specified groups that do not capture the diversity of the human experience. In other medical specialties, race correction factors have historically been incorporated into medical calculations, such as for estimating lung capacity in pulmonology and estimating kidney function in nephrology. These factors have not only been disputed and discredited for being founded on inaccurate assumptions about race, but have resulted in worse outcomes and widening health disparities for African Americans. Race and ethnicity, as social constructs, reflect a complex mix of genetic, socioeconomic, political, and cultural factors. These result in heterogeneous phenotypic presentations where inter-group diversity often exceeds intra-group diversity. Ignoring this variability compromises the diagnostic ability of tests like the OCT, underscoring the need for a more holistic, nuanced approach to developing normative databases.

Instead of limiting the number of individuals included in our normative databases to those who meet strict inclusion criteria, we need to focus on expanding our normative datasets to be more broadly representative of those who will use the test. To be clinically useful as a screening tool, the OCT needs to have high sensitivity and specificity and the confidence interval around classifying an optic nerve as “red” needs to be very narrow. Currently, for example, in the Topcon Maestro normative dataset, the 1% (red) threshold is based on values from 9 eyes and has a confidence interval of 0.4% to 2.5%.¹³ While this sample size was chosen to ensure that the confidence intervals between the different color codes did not overlap, the width of the confidence interval was not pre-specified as part of the power calculation in estimating the sample size needed to create a normative database.¹³ Alternately, to have a truly narrow confidence interval where no fewer than 0.9% and no more than 1.1% of healthy people would be classified as “red” – where “red” is supposed to identify healthy participants whose measured RNFL thickness is the lowest 1 percentile - a sample size of 50,000 patients would be needed in the normative database (Supplemental Table 2).

As we move towards the use of automated analyses of ophthalmic imaging, training and validation of algorithms on multitudes of heterogeneous participant data is essential. In a global review of publicly available datasets that include ophthalmological imaging for the development of machine learning algorithms to detect ocular disease, 18 of 94 datasets included OCT.⁴⁶ Of these, datasets that included both healthy eyes and eyes with ocular diseases from anywhere between 10 to 5319 individuals were included.⁴⁶ To obtain really large normative datasets, on the order of 50,000 participants, we will need to leverage large national and multinational registries such as The American Academy of Ophthalmology IRIS^® Registry (Intelligent Research in Sight)⁴⁷, the Sight Outcomes Research Collaborative (SOURCE) registry⁴⁸, the IMVISUAL consortium⁴⁹, the UK BioBank⁵⁰, the Million Veteran Program⁵¹, and the All of Us Research program⁵². Currently the IRIS registry, UK BioBank, SOURCE, IMVISUAL, and All of Us all include ophthalmic imaging. Such large datasets inform biomedical research and intervention by enabling clinicians and researchers to generalize findings across populations with greater confidence, ensuring that a wider range of individual differences are represented, and providing greater statistical power to detect significant differences. At the very least, further transparency about the characteristics of a reference database is imperative for informed clinical decision making. Clinicians also need to be aware of the limitations of making predictions based on an overreliance of the quadrant maps. Rather, careful scrutiny of B-scans for segmentation defects and thin, focal areas of loss, as well as longitudinal changes within an individual is crucial.

There are several limitations to this study. First, all healthy subjects in the Topcon normative database were required to have a normal visual field. Healthy individuals attending our community screening program, on the other hand, were not obligated to undergo an additional visual field test. Other studies comparing color code labels in healthy cohorts also did not necessarily require visual function testing and found similar discrepancies in the actual versus expected color code distributions.^21,24,30,40 While having a visual field for all participants would be ideal, this is not practical in a community screening setting. In addition, because all glaucoma assessments were made without a visual field test, we purposefully included participants who had abnormal OCT if they met no other criteria for glaucoma based on a tele-ophthalmology model of screening for eye disease (normal fundus, normotensive intraocular pressure, no self-reported diagnosis of glaucoma) and therefore may have included individuals who were otherwise considered glaucoma suspects or had pre-perimetric disease. Second, we broadly categorized individuals as White or Black/African American based on self-report from initial intake forms. As acknowledged previously, this is a heterogenous category. We also did not include individual self-identifying as other races due to small sample sizes. Third, we could not differentiate RNFL thickness measurements with white labels due to thicknesses in the top 95^th percentile from those with white labels due to undefined age or disc area. Finally, we did not investigate the suspected interaction effects between the various ocular parameters and sociodemographic factors in predicting RNFL color labels as our study was not powered to do so.

In this study, we compared OCT parameters between the Topcon normative cohort and a large, diverse, real-world cohort of healthy participants from the MI-SIGHT program. Although there were no differences in the mean ONH parameters and retinal thickness measures, significant differences emerged when comparing the Topcon normative cohort to the MI-SIGHT healthy cohort when stratified by self-identified race. Furthermore, the distribution of RFNL thickness color labels in the MI-SIGHT cohort deviated from the expected distributions, despite being a healthy cohort. While color label distributions across various ethnically homogeneous cohorts have been previously studied, this study explores the application of the OCT normative database in a real-world community setting in both White and Black/African American individuals. Additionally, we used a multivariable model to gain deeper insights into predictors of color label classifications. Our findings reveal that factors influencing color labels extend beyond the commonly cited influences of self-identified race and disc area. Our study’s results support the argument that to accurately differentiate between normal and glaucomatous eyes in real-world populations with OCT, larger datasets that include individuals with a broad range of ocular parameters and sociodemographic backgrounds are necessary. Only then can the full potential of OCT be utilized for screening, particularly in communities at high risk for blindness from glaucoma.