Evaluation of the Cirrus HD-OCT Normative Database Probability Codes in an Black American Population

Abstract

Objective:

Race-adjusted interpretation of data from Cirrus high-definition optical coherence tomography (HD-OCT) devices is not standard practice. The aim of this study is to evaluate differences in peripapillary retinal nerve fiber layer (RNFL) thickness between healthy Black Americans and the Cirrus HD-OCT normative database.

Design:

This is a cross-sectional observational study using control patients recruited from the greater Philadelphia, Pennsylvania area.

Subjects:

A total of 466 eyes were included in this study. Subjects were retrospectively identified from the control cohort of the Primary Open-Angle Black American Glaucoma Genetics (POAAGG) study.

Methods:

Using an age-stratified or linear regression method, we reclassified white-green-yellow-red color probability codes for RNFL thicknesses by quadrant.

Main Outcomes Measures:

The distribution of reclassified color codes was compared to the expected 5%-90%-4%-1% percentiles and to the original color codes by the Cirrus machine.

Results:

Average RNFL thickness in the POAAGG control cohort was thinner than the Cirrus normative database in all except the nasal quadrant. The original color codes of the POAAGG cohort did not fall into the expected distributions, with more RNFL measurements assigned as white and red codes than expected (9.5% and 1.7%) and fewer measurements assigned as green and yellow codes than expected (85.3% and 3.5%) (p<0.001). Compared to the original Cirrus machine, reclassification using linear regression produced color codes closest to the expected distributions (p=0.09). The proportion of abnormal results shifted closer to the expected 5% in the nasal (1.3%, p<0.001 vs. 3.0%, p=0.048) and temporal (8.2%, p=0.002 vs. 3.6%, p=0.18) quadrants.

Conclusions:

Results further establish the presence of structural differences in the RNFL of Black American patients. Color code reclassification suggests that the existing Cirrus database may not be accurately evaluating glaucomatous nerves in patients of African descent. This study addresses an unmet need to assess Cirrus HD-OCT color probability codes in a Black American population.

Précis

An alternative set of color probability code cutpoints for abnormality yielded values closer to the expected distributions for healthy eyes of Black Americans compared to the Cirrus HD-OCT normative database.

Introduction

Primary open-angle glaucoma (POAG) is a progressive optic neuropathy that is the leading cause of irreversible blindness in developed countries^{1, 2}. POAG is more common, more refractory to treatment, and more severe in Black Americans compared to other ethnic groups³. Though its etiology is multi-factorial and not fully understood, the disease is known to be characterized by a slow degeneration of retinal ganglion cells⁴. The loss of these retinal cells can result in an abnormal appearance of the optic nerve head (ONH), with corresponding loss of visual field⁵.

Early diagnosis and detection of POAG is essential to prevent irreversible vision loss. One of the main tools used for diagnosis is evaluation of abnormalities in the ONH and retinal nerve fiber layer (RNFL). Spectral domain Cirrus high-definition-optical coherence tomography (HD-OCT; Carl Zeiss Meditec, Dublin, CA) is a non-invasive imaging technique that can be used to evaluate RNFL and ONH parameters. Numerous studies have demonstrated that it has high sensitivity and specificity in discriminating between healthy and glaucomatous eyes, and it is commonly used to diagnose glaucoma and monitor its progression in clinical practice^{6, 7}.

The Cirrus HD-OCT software compares patient data to a normative database, using a proprietary algorithm to determine whether patient values are within normal limits. Currently, all ONH parameter limits are specific to age and disc area, while RNFL thickness limits are specific to age only. Each patient value is assigned a color probability code corresponding to where it is located in the distribution of thresholds within the normative database [white: above the 95^th percentile (5%); green: within the 95^th and 5^th percentiles (90%); yellow: within the 1^st and 5^th percentiles (4%); red: at or below the 1^st percentile (1%)], with yellow and red indicating abnormal values. One factor that limits the ability of HD-OCT to detect glaucoma is the large variability of ONH parameters among healthy individuals⁸. Furthermore, several studies using multiple imaging modalities have shown that racial differences exist in RNFL and ONH parameter measurements^{3, 9-23}. Given these racial differences, it is important to examine whether the normative database is accurate for classifying high-risk minority populations. The current RNFL normative database for the Cirrus HD-OCT contains data from 282 healthy subjects aged 19 to 84, of which 51 (18%) are Black American. The average age of patients in the normative database is 46.5 years, and it includes only three patients over the age of 80^{24, 25}.

The purpose of this study was to investigate differences in RNFL thickness between Cirrus HD-OCT normative database subjects and a cohort of Black American patients without glaucoma. These patients were previously enrolled in the Primary Open-Angle Black American Glaucoma Genetics (POAAGG) study, which seeks to elucidate the genetic architecture of POAG in Black Americans²⁶. A RNFL normative database specific to Black American patients may be more useful in detecting and monitoring glaucoma in this overaffected population.

Methods

Participants

The POAAGG study population consists of self-identified Black Americans and individuals of African or Afro-Caribbean descent, over age 35, recruited from the greater Philadelphia region. Fellowship-trained glaucoma specialists determined each subject’s classification as a glaucoma case, control, or suspect based on previously published criteria²⁶. Controls in the POAAGG study were identified as Black American patients without (1) high myopia ( > −8.00 diopters); (2) high hyperopia (> +8.00 diopters); (3) family history of POAG; (4) ocular hypertension (intraocular pressure >21 mmHg); (5) evidence of neuroretinal rim thinning, excavation , notching, nerve fiver layer defects, optic nerve asymmetry, or a cup-to-disc ratio difference greater than 0.2 between eyes; or (6) significant ocular pathology including iritis, uveitis, or iridocyclitis; Graves’ disease with ocular manifestations; vascular occlusion causing iris neovascularization; and optic nerve atrophy from other diagnoses. The majority of study participants in the POAAGG control cohort did not have a Humphrey 24-2 SITA standard visual field. The University of Pennsylvania Institutional Review Board approved this study and the informed consent process. All participants provided informed consent.

Inclusion and Exclusion Criteria

Black American patients were retrospectively identified from control subjects enrolled in the POAAGG study²⁶. Patients that met the basic inclusion criteria employed in the Cirrus HD-OCT database study²⁴ were included in the present study. Patients with diabetes and hypertension were included in the present study, though patients with history of proliferative diabetic retinopathy, diabetic macular edema, and retinal vascular occlusion were excluded.

The spectral domain OCT device used for all subjects was Cirrus HD-OCT (Model 5000; Carl Zeiss Meditec Inc., Dublin, CA). For patients with multiple OCT scans taken on the date of enrollment in the POAAGG study, the scan with the highest signal strength (six or higher) was used. Scans of signal strength five or lower were excluded. If there were multiple scans with the same highest signal strength for a given eye, the first scan in the clinical database was selected for data collection. The scan signal strength, RNFL thickness, RNFL symmetry, rim area, disc area, average cup to disc (C/D) ratio, vertical C/D ratio, and cup volume were collected. The color probability code was noted for each ONH parameter, RNFL quadrant, and clock hour.

A thorough chart review for each POAAGG control subject was completed. Diagnoses of diabetes, hypertension, previous cataract surgery, amblyopia, leukemia, AIDS, dementia, and multiple sclerosis were noted. The clinical data of each subject was subsequently reviewed by a glaucoma specialist. Patients were excluded for poor quality scans as well as diagnoses of proliferative diabetic retinopathy, macular edema, retinal occlusion, amblyopia, and systemic conditions including leukemia, AIDS, dementia, and multiple sclerosis. Patients with prior cataract surgery were not excluded from this study.

Statistical Analysis

Characteristics were compared using a two-sample t-test based on the mean, standard deviations (SDs), and sample sizes calculated for the POAAGG cohort, and those reported by Knight et al²⁴ for the Cirrus normative database. A reduced "effective sample size" was calculated as the existing sample size divided by (1 plus the intraclass correlation of each measure), and used in place of the actual number of eyes for the POAAGG controls to account for the correlation between eyes²⁷.

Means and SDs of RNFL values of the POAAGG controls from all four quadrants and for each quadrant separately were calculated for each age decade, and ranges of the observed RNFL measures were shown within these strata for each color code generated by the Cirrus HD-OCT. For each age strata, adjusted cutpoints between color codes based on the POAAGG control data were calculated by percentile of the normal distribution, using the standard deviation, as red to yellow (1^st percentile), yellow to green (5^th percentle), and green to white (95^th percentile). As a second approach to determining the cutpoints, we used linear regression models of the RNFL, including age as a continuous covariate and stratified by quadrant, to estimate the normative limits of RNFL thickness. For each quadrant, the standardized residuals were classified by percentile of the normal distribution to determine the cutpoints, such that values greater than the 95th percentile (about 1.64) were classified as white, values less than white but greater than the 5th percentile (about −1.64) were green, values less than green but greater than the 1st percentile (about −2.33) were yellow, and values less than yellow were red.

A chi-square test was used to test the goodness of fit of the classifications by the Cirrus HD-OCT, by the age-strata method, and by the linear regression method, to the expected 1% red, 4% yellow, 90% green, and 5% white. Weighed kappa statistics were calculated to assess the agreement of the three methods for classifying RNFL values into one of the four color probability codes. All analyses were performed with SAS version 9.4 (Cary, NC).

Results

Demographics and ocular characteristics:

A total of 466 eyes from the POAAGG control group were included for final analysis. Table 1 presents the demographic and ocular characteristics of the POAAGG cohort. Subjects in the control group were older on average than subjects in the Cirrus RNFL normative database²⁴ (59.1 vs. 46.2 years, p <0.001). A total of 104 eyes from patients 70 years or older were included in the POAAGG control cohort. Subjects were majority female (67.4%), and there was an equal distribution of eye laterality. A total of 36.1% and 54.3% of POAAGG controls had a co-morbid diagnosis of diabetes and hypertension, respectively. Averages of the maximum intraocular pressure and clinically determined cup-to-disc (C/D) ratio were within normal limits (17.0 mmHg; C/D ratio 0.34).

Table 1.

Demographic and ocular characteristics of 466 healthy Black American eyes from the Primary Open-Angle Black American Glaucoma Genetics (POAAGG) study.

	POAAGG Controls	Cirrus HD-OCT RNFL Normative Database25
Sample Size, eyes	466	282
Age, years, mean (SD; range)	59.1 (12.1; 33-87)	46.2 (16.9; 19-84)
Gender, male:female	152 : 314	128 :143
Laterality, OD:OS	233 : 233	-
History of diabetes, eyes, %	168 (36.1%)	-
History of hypertension, eyes, %	253 (54.3%)	-
Intraocular pressure, mmHg, SD; eyes	17.0 (3.2) ; n=430a	14.1 (2.4)
Clinically-determined cup-to-disc ratio, SD, eyes	0.34 (0.13) ; n=392	-

SD: standard deviation

^aFor POAAGG controls, the maximum intraocular pressure was used for analysis.

Comparing optic nerve parameters:

Table 2 compares ONH parameters and RNFL thickness between the POAAGG control cohort and the Cirrus RNFL normative database. Average signal strength of patient OCT data in the control cohort was 8.29. The average rim area (1.41 vs. 1.31 mm², p <0.001), average C/D ratio (0.51 vs. 0.46, p <0.001), vertical C/D ratio (0.49 vs. 0.44, p <0.001), and cup volume (0.17 vs. 0.14 mm³, p=0.032) were all slightly higher in the POAAGG study controls than in the Cirrus database patients^{24, 25}. Among the ONH parameters, the greatest difference between the groups was the average disc area. In the Cirrus database subjects, the minimum disc area was 1.00 mm², the maximum disc area was 2.93 mm², and the average disc area was 1.77 mm² (SD 0.34 mm²)^{24, 25}. In comparison, amongst the POAAGG control subjects, the minimum disc area was 0.88 mm², the maximum disc area was 3.73 mm², and the average disc area was 2.01 mm² (SD 0.40 mm², p <0.001). However, the difference in average disc area diminishes if compared to the average disc area adjusted for age in those of African descent in the Cirrus normative database (1.93 mm²)²⁴ (data not shown in Table 2).

Table 2.

Optic nerve head parameters and RNFL thicknesses of 466 eyes of healthy Black Americans compared to the eyes used to develop the Cirrus HD-OCT RNFL normative database.

	POAAGG Controls	Cirrus HD-OCT RNFL Normative Database25	P value
Sample Size, eyes	466	282	-
Signal Strength, mean, SD	8.29 (1.09)	-	-
Disc Area, mm2, SD
Average	2.01 (0.40)	1.77 (0.34)	<0.001
Minimum	0.88	1.00	-
Maximum	3.73	2.93	-
Rim Area, mm2, SD	1.41 (0.26)	1.31 (0.22)	<0.001
Average CDR, mean, SD	0.51 (0.15)	0.46 (0.17)	<0.001
Vertical CDR, mean, SD	0.49 (0.14)	0.44 (0.17)	<0.001
Cup Volume, mm3, SD	0.17 (0.14)	0.14 (0.13)	0.002
RNFL Thickness, μm, SD
Average	92.9 (10.63)	94.0 a	-
Inferior	122.1 (29.34)	123.2 a	-
Nasal	73.8 (19.72)	69.8a	-
Superior	116.6 (25.20)	119.0 a	-
Temporal	59.2 (14.52)	64.0 a	-

RNFL: retinal nerve fiber layer; CDR: cup-to-disc ratio; SD: standard deviation

^aRNFL thickness value adjusted for age and disc area²³

Average RNFL thickness in the POAAGG cohort was 92.9 μm, similar to the adjusted RNFL thickness for age and disc area among the Cirrus database patients (94.0 μm)²⁴, but thicker than those of only European-descent (90.1 μm) (data not shown in Table 2). All individual quadrants in the POAAGG cohort were thinner than the adjusted RNFL thickness of the Cirrus database patients (inferior: 122.1 vs 123.2 μm; superior: 116.6 vs 119.0 μm; temporal: 59.2 vs 64.0 μm), except for the nasal quadrant (73.8 vs 69.8 μm) (Table 2).

We examined the distribution of RNFL thickness by quadrant in the POAAGG control group (Table 3). Considering all four quadrants and the average RNFL thickness for each eye, the RNFL values were not coded by the Cirrus machine according to the expected 4-color distributions, with white (9.5%) and red (1.7%) being over-represented and green (85.3%) and yellow (3.5%) being under-represented (p<0.001). The nasal quadrant produced fewer than expected abnormal results (0.9% yellow, 0.4% red, 1.3% total, p<0.001), and the temporal quadrant produced more than expected abnormal results (4.9% yellow, 3.2% red, 8.2% total, p=0.002). The inferior and superior quadrants as well as average RNFL thickness produced roughly expected results (p=0.78, p=0.32, and p = 0.57, respectively).

Table 3:

Expected and observed color codes, by method, and their agreement with the original by quadrant

			Color (expected %)						Combined color (expected %)
Quadrant	Method	n	White (5%)	Green (90%)	Yellow (4%)	Red (1%)	P value*	Weighted Kappa**	White/Green (95%)	Yellow/Red (5%)	P value*	Kappa**
All	Machine	2330	9.5%	85.3%	3.5%	1.7%	<0.001	.	94.8%	5.2%	0.74	.
	Age-strata	2330	5.3%	91.2%	2.9%	0.6%	0.007	0.60	96.5%	3.5%	<0.001	0.70
	Linear	2330	5.4%	90.7%	3.2%	0.7%	0.09	0.65	96.1%	3.9%	0.02	0.77
Average	Machine	466	13.1%	81.3%	4.5%	1.1%	<0.001	-	94.4%	5.6%	0.57	-
	Age-strata	466	4.3%	91.4%	3.2%	1.1%	0.73	0.61	95.7%	4.3%	0.48	0.82
	Linear	466	4.5%	90.6%	4.3%	0.6%	0.82	0.66	95.1%	4.9%	0.95	0.94
Inferior	Machine	466	12.9%	82.4%	3.6%	1.1%	<0.001	.	95.3%	4.7%	0.78	.
	Age-strata	466	4.7%	91.8%	3.0%	0.4%	0.40	0.56	96.6%	3.4%	0.12	0.73
	Linear	466	5.6%	90.8%	3.2%	0.4%	0.46	0.63	96.4%	3.6%	0.18	0.81
Nasal	Machine	466	12.4%	86.3%	0.9%	0.4%	<0.001	.	98.7%	1.3%	<0.001	.
	Age-strata	466	5.8%	91.6%	2.6%	0.0%	0.21	0.60	97.4%	2.6%	0.02	0.66
	Linear	466	5.8%	91.2%	2.6%	0.4%	0.21	0.62	97.0%	3.0%	0.048	0.59
Superior	Machine	466	8.4%	85.6%	3.4%	2.6%	<0.001	.	94.0%	6.0%	0.32	.
	Age-strata	466	6.7%	89.9%	2.4%	1.1%	0.13	0.78	96.6%	3.4%	0.12	0.71
	Linear	466	6.4%	89.3%	2.8%	1.5%	0.18	0.83	95.7%	4.3%	0.48	0.82
Temporal	Machine	466	0.9%	91.0%	4.9%	3.2%	<0.001	.	91.8%	8.2%	0.002	.
	Age-strata	466	4.9%	91.4%	3.4%	0.2%	0.34	0.43	96.4%	3.6%	0.18	0.60
	Linear	466	4.5%	91.8%	3.2%	0.4%	0.45	0.46	96.4%	3.6%	0.18	0.60

^*P-values represent an overall difference from the expected percentiles (5%/90%/4%/1% for color, 95%/5% for combined color)

^**Kappa values represent the agreement of the classification colors by the machine method with those generated by the age-strata method, and by the linear regression method

Identifying new cutpoints and reclassification:

To identify new color code cutpoints for the Black American POAAGG control cohort (shown in Table 4), we used a previously characterized method that stratifies by age and identifies cutoff points based on the standard deviation^{23, 28, 29}. Considering all RNFL values, the resulting color codes did not fall into the expected distributions (white 5.3%, green 91.2%, yellow 2.9%, and red 0.6%, p=0.007, Table 3). Compared to the Cirrus machine, the number of abnormal results (yellow and red codes) shifted further from the expected result (5%) for all RNFL thicknesses (3.5%, p<0.001, kappa=0.70), but became closer to expected for the nasal (2.6%, p=0.02, kappa=0.66) and temporal (3.6%, p=0.18, kappa=0.60) quadrants.

Table 4.

Age-stratified RNFL values and reclassified cutpoints by quadrant

	Total				Red Code (0-1%)	Yellow Code (1-5%)	Green Code (5-95%)	White Code (95-100%)
Quadrant	Age (yrs)	N Obs	Mean (μm)	Std. Dev	Adjusted Cutpoints	Adjusted Cutpoints	Adjusted Cutpoints	Adjusted Cutpoints
All	30-39	100	97.5	29.9	< 28.1	28.1 – 48.4	48.4 – 146.6	> 146.6
	40-49	510	94.6	28.4	< 28.6	28.6 – 47.9	47.9 – 141.2	> 141.2
	50-59	610	93.8	28.7	< 27.1	27.1 – 46.6	46.6 – 141.0	> 141.0
	60-69	590	92.3	27.8	< 27.6	27.6 – 46.5	46.5 – 138.0	> 138.0
	70+	520	90.1	26.8	< 27.8	27.8 – 46.0	46.0 – 134.2	> 134.2
Average	30-39	20	97.6	11.2	< 71.5	71.5 – 79.2	79.2 - 116.0	> 116.0
	40-49	102	94.6	9.7	< 72.0	72.0 – 78.6	78.6 - 110.5	> 110.5
	50-59	122	93.8	9.5	< 71.7	71.7 – 78.2	78.2 - 109.4	> 109.4
	60-69	118	92.2	11.5	< 65.6	65.6 – 73.4	73.4 - 111.1	> 111.1
	70+	104	90.1	11.0	< 64.4	64.4 – 72.0	72.0 - 108.3	> 108.3
Inferior	30-39	20	128.3	19.9	< 82.0	82.0 – 95.5	95.5 – 138.6	> 161.0
	40-49	102	122.6	18.1	< 80.6	80.6 – 92.9	92.9 – 161.0	> 152.4
	50-59	122	125.4	16.8	< 86.2	86.2 – 97.7	97.7 – 153.1	> 153.1
	60-69	118	121.9	19.3	< 76.9	76.9 – 90.1	90.1 – 153.7	> 153.7
	70+	104	116.9	19.6	< 71.4	71.4 – 84.7	84.7 – 149.0	> 149.0
Nasal	30-39	20	74.4	14.7	< 40.0	40.0 – 50.1	50.1 – 98.6	> 98.6
	40-49	102	74.7	10.5	< 50.4	50.4 – 57.5	57.5 – 92.0	> 92.0
	50-59	122	73.3	13.3	< 42.3	42.3 – 51.4	51.4 – 95.3	> 95.3
	60-69	118	75.1	13.6	< 43.6	43.6 – 52.8	52.8 – 97.5	> 97.5
	70+	104	71.6	13.4	< 40.6	40.6 – 49.7	49.7 – 93.6	> 93.6
Superior	30-39	20	124.6	14.4	< 91.1	91.1 – 100.9	100.9 – 93.6	> 148.2
	40-49	102	120.6	19.0	< 76.5	76.5 – 89.5	89.5 – 151.8	> 151.8
	50-59	122	117.9	15.9	< 80.8	80.8 – 91.7	91.7 – 144.1	> 144.1
	60-69	118	113.7	17.9	< 72.1	72.1 – 84.3	84.3 – 143.1	> 143.1
	70+	104	112.9	15.8	< 76.1	76.1 – 86.9	86.9 – 138.8	> 138.8
Temporal	30-39	20	62.8	10.1	< 39.3	39.3 – 46.2	46.2 – 79.5	> 79.5
	40-49	102	60.3	9.8	< 37.4	37.4 – 44.1	44.1 – 76.4	> 76.4
	50-59	122	58.7	9.1	< 37.6	37.6 – 43.8	43.8 – 73.6	> 73.6
	60-69	118	58.4	9.4	< 36.5	36.5 – 42.9	42.9 – 73.9	> 73.9
	70+	104	59.1	11.3	< 32.8	32.8 – 40.5	40.5 – 77.7	> 77.7

The age-stratified method created unexpected results due to low sample size within each age strata. For example, the adjusted cutpoint between yellow and green (i.e. between normal and abnormal) for the nasal quadrant fluctuated between 49.7 and 57.5 μm (Table 4) when a gradual decrease in thickness is expected with age. To make the reclassification more robust, we reclassified all of the colors by fitting a linear regression of RNFL thickness by continuous age, stratified by quadrant, then categorizing the standardized residuals into the expected distribution of colors. This method is most similar to the “fitted regression” methodology²⁵ that Cirrus HD-OCT uses for color codes in which by age, residuals were calculated for each model and used to estimate the 1^st, 5^th, 95^th, and 99^th percentiles of normal distribution. Figure 1 graphically demonstrates the RNFL values and reclassified cutpoints using this method by quadrant, which follow the expected downward trend of RNFL thickness with age. This produced results (Table 3) that were closest to the expected distribution (white 5.4%, green 90.7%, yellow 3.2%, and red 0.7%, p=0.09). Compared to the Cirrus machine, the number of abnormal results shifted further from the expected result for all RNFL thicknesses (3.9%, p<0.02, kappa=0.77) but became closer to expected for the nasal (3.0%, p=0.048, kappa=0.59), and temporal (3.6%, p=0.18, kappa=0.60) quadrants.

Figure 1:

Retinal nerve fiber layer (RNFL) thickness values and reclassified cutpoint ranges using the linear regression method by quadrant and average. Red triangles indicate RNFL values that changed color codes upon reclassification.

Discussion

The purpose of this study was to evaluate differences in RNFL measurements in a healthy Black American population compared to the Cirrus HD-OCT normative database. Despite studies suggesting differences in the optic nerve parameters based on ethnicity, it is not standard practice to account for race when determining normal ranges of OCT measurements³⁰. We used two alternative methods of establishing cutpoints and found that the linear regression method produced results closest to the expected color code distributions. This study addresses an unmet need to evaluate Cirrus HD-OCT color probability codes in a Black American population.

Overall, RNFL thickness in this healthy POAAGG cohort followed the ISNT rule of descending quadrant thickness (inferior followed by superior, nasal, and temporal). Compared to patients of European descent in the Cirrus normative database²⁴, all quadrants except for the temporal region were thicker in the POAAGG controls. These results correspond to existing literature^{13, 22, 24, 31, 32} and are in agreement with the African Descent and Glaucoma Evaluation Study, which similarly noted greater overall RNFL thickness, with an exception over the papillomacular bundle, in patients of African versus European descent¹³. Notably, in comparison to the entire Cirrus cohort, POAAGG controls had thinner RNFL in all but the nasal quadrant. This is likely due to the fact that Chinese and Hispanic patients, who make up 36% of the Cirrus cohort²⁵, had significantly thicker RNFL in all quadrants, raising the average RNFL thickness in the overall cohort.

The current Cirrus algorithm does not account for this variation in RNFL thickness across ethnicities, limiting the ability to properly discrimate between normal and abnormal optic nerve measurements. In this study, we show how reclassification of the color codes in an Black American population produces color code distributions closest to the expected percentiles, particularly in the temporal and nasal quadrants. Because the RNFL of the temporal quadrant in this Black American cohort is thinner than the overall Cirrus database, more RNFL values are coded as abnormal (8.2%) than expected (5%) by the Cirrus machine. This improves upon reclassification with the linear regression method (3.6%). The opposite effect is seen in the nasal quadrant, with fewer RNFL values coded as abnormal (1.3%) than expected by the Cirrus machine and subsequent improvement upon reclassification (3.0%). These findings suggest that in the temporal quadrant, existing RNFL thicknesses coded by the Cirrus machine are likely to overdiagnose abnormal optic nerves in Black Americans whereas in the nasal quadrant, abnormal thinning may be more likely to be coded as normal by the Cirrus machine.

There is growing interest in the literature for establishing group-specific normative databases to improve the sensitivity or specificity of OCT color probability codes³⁰. Multiple papers have evaluated control OCT databases for Asian populations^{23, 28, 33-36}, with some noting inconsistences between the distribution of RNFL thicknesses of their respective populations and Cirrus normative databases. Hong et al²⁸ found that once color codes were adjusted based on a healthy Korean cohort, 4% of average RNFL measurements that were originally coded as normal by Cirrus HD-OCT were reclassified as abnormal. Similarly, a study utilizing a healthy Vietnamese population²³ reported a statistically significant increase of red and yellow codes in the average and inferior peripapillary RNFL, following color code adjustment compared to the original Cirrus HD-OCT normative database. In addition to normative databases targeting ethnicity, several studies have shown that a myopic normative database can improve the specificity of detecting glaucoma in myopic eyes^{29, 37-39}. Because the RNFL tends to be thinner in myopic eyes^{40, 41}, abnormal color codes that reclassify as normal may be more prevelant in this population. In a cohort of myopic Korean eyes with glaucoma, for example, Seol et al²⁹ showed poor agreement between the built-in spectral domain OCT and myopic normative databases and demonstrated an increase in diagnostic specificity after re-evaluating with the myopic normative database. Overall, group-specific normative databases may allow for a more accurate interpretation of ONH data.

There are several limitations to this study. First, while the Cirrus normative database excluded patients with diabetes, we were unable to exclude all subjects with diabetes from our cohort due to the prevalence of the disease within the population. We did, however, exclude all patients with proliferative diabetic retinopathy and/or macular edema. Given reported associations between diabetes and RNFL loss⁴², we further compared RNFL thicknesses in POAAGG subjects with and without diabetes and found only small differences between groups. While the inferior quadrant was slightly thinner in diabetics (−5.2 μm), the other three quadrants did not significantly differ between groups. This provided further assurance that the inclusion of diabetic patients in our cohort made little difference in the final analysis. Similarly, all subjects in the Cirrus normative database were required to have a normal visual field, whereas a normal visual field was not required for our POAAGG control cohort. Additionally, data from the POAAGG study is limited to the greater Philadelphia community and may not be reflective of all people of African descent. However, an attempt at inclusivity was made by enrolling a cohort that includes subjects who self-identified as Black American, African descent, and Afro-Caribbean. Finally, because Cirrus HD-OCT software calculates cutpoints for the color probability codes based on a propriatory software to which we do not have access, we were unable to directly replicate how an HD-OCT device would interpret data from an Black American normative database. However, compared to the age-stratified method, the linear regression model used in this paper produced color codes distributions closest to the expected percentiles and is most similar to the “fitted regression” methodology²⁵ that Cirrus HD-OCT uses to determine color codes. The fact that reclassified color codes did not adhere to the expected 5% distribution of abnormal results is most likely due to a small sample size, but a non-linear relationship with age, and/or a non-gaussian distribution of normal RNFL thicknesses, may also be contributors.

In conclusion, these findings emphasize the presence of structural differences in the RNFL of Black American patients. Using a linear regression model, we reclassified the color probability codes and demonstrated an improved rate of identifying abnormality in the nasal and temporal quadrants. Ultimately, the color code reclassification generated based on the POAAGG control cohort will need be evaluated against a cohort of glaucoma cases to determine whether it leads to increased sensitivity and/or specificity for detecting disease.

Abbreviations/Acronyms:

HD-OCT

high-definition optical coherence tomography

POAAGG

Primary Open-Angle Black American Glaucoma Genetics

RNFL

retinal nerve fiber layer

POAG

Primary open-angle glaucoma

ONH

optic nerve head

SD

standard deviation

C/D

cup-to-disc