Influence of Diabetes and Diabetic Retinopathy on the Performance of Heidelberg Retina Tomography II for Diagnosis of Glaucoma

Abstract

Purpose.: To determine whether diabetes and diabetic retinopathy (DR) affect the performance of the Heidelberg Retina Tomograph II (HRT II; Heidelberg Engineering, Heidelberg, Germany) algorithms for glaucoma detection.

Methods.: This population-based survey was conducted among Malays in Singapore who were a 40 to 80 years of age. Diabetes was defined as self-report of a physician's diagnosis, use of diabetic medication, or a random blood glucose level ≥11.1 mmol/L. Retinal photographs were graded for DR according to the modified Airlie House classification system. The diagnosis of glaucoma was based on International Society for Geographical and Epidemiologic Ophthalmology criteria. The sensitivity and the false-positive rates were calculated for the Moorfields regression analysis [MRA]; the linear discriminant functions (LDFs) by Mikelberg (Mikelberg-LDF), Burk (Burk-LDF), and Bathija (Bathija-LDF); and the support vector machine (SVM).

Results.: A total of 1987 persons without diabetes (including 86 with glaucoma) and 524 with diabetes (including 26 with glaucoma) were analyzed. The presence of diabetes had no influence on both the sensitivities and false-positive rates for all HRT algorithms. In the multivariate analyses adjusting for optic disc size, the presence of DR was significantly associated with the higher false-positive rates for Burk-LDF and Bathija-LDF (P < 0.05), but not with the false-positive rates for MRA, Mikelberg-LDF, and SVM.

Conclusions.: Diabetes does not affect the performance of HRT II for diagnosis of glaucoma, but the presence of DR may be a source of false-positive test results.

The Heidelberg Retina Tomograph II (HRT II, Heidelberg Engineering GmbH; Heidelberg, Germany) is an instrument used increasingly for the assessment of the optic disc for the diagnosis of glaucoma. ^1,2 Studies have been conducted to examine factors that may affect the performance of HRT. Age, optic disc size, glaucoma disease severity, and refractive error have been identified as possibly affecting the diagnostic accuracy of HRT II or HRT, version 3.0. ^{3 –10}  

However, little is known about the influence of diabetes, which affects up to 10% of the adult population in many countries. ^11,12 Although clinic-based reports have suggested that persons with diabetes or diabetic retinopathy (DR) are more likely to have a thinner retinal nerve fiber layer (RNFL), ^13,14 there is no report of the impact of diabetes or DR on the performance (sensitivity and specificity) of HRT algorithms for diagnosing glaucoma in a population-based setting. Such information is clinically important, as many people with diabetes also have glaucoma. ^15,16  

In this study, we examined the influences of diabetes and DR on the sensitivity and specificity of HRT algorithms for diagnosing glaucoma in a population-based study of Asian Malays. 

Methods

Study Population

The Singapore Malay Eye study (SiMES) was a population-based, cross-sectional study of adult Malays living in Singapore. ^17,18 It was conducted in accordance with the World Medical Association Declaration of Helsinki. Ethics approval was obtained from the Singapore Eye Research Institute (SERI) Institutional Review Board. Details of the study have been reported elsewhere. ^17,18 In brief, an age-stratified random sampling procedure was used to select persons of Malay ethnicity aged 40 to 80 years living in the southwestern part of Singapore. A questionnaire was administrated to collect demographic and medical history including the known duration of diabetes. Nonfasting venous blood samples were obtained and used to assess serum blood glucose and glycosylated hemoglobin (HbA1c). Diabetes mellitus was defined as self-reported physician's diagnosis, use of diabetic medication, or nonfasting blood glucose ≥11.1 mmol/L. ^19,20  

Measurement and Definition of Glaucoma

The measurement and definition of glaucoma have been described elsewhere. ^{21 –23} Intraocular pressure (IOP) was measured with Goldmann applanation tonometry (Haag-Streit, Köniz, Switzerland) before pupil dilation. Automated perimetry (SITA FAST 24-2, Humphrey Visual Field Analyzer II; Carl Zeiss Meditec, Dublin, CA) was performed with near refractive correction on: (1) one in five consecutive participants without suspected glaucoma (n = 641 persons) before ocular examination and (2) all participants with suspected glaucoma (defined later). The visual field test was repeated on another occasion if the test reliability criteria were not satisfied (fixation losses >20%, false positives >33%, and/or false negatives >33%) or if there was a visual field defect (defined later). After pupil dilation, the optic disc was evaluated and the vertical cup-to-disc ratio (VCDR) was determined with a +78-D lens, at ×16 magnification. ^22,23  

All patients with suspected glaucoma underwent visual field testing and were defined by any of the following criteria: (1) IOP >21 mm Hg; (2) VCDR >0.6 or VCDR asymmetry >0.2; (3) abnormal anterior segment deposit consistent with pseudoexfoliation or pigment dispersion syndrome; (4) occludable angle, defined as posterior trabecular meshwork seen for ≤180° of the angle circumference during static gonioscopy; (5) peripheral anterior synechiae or other findings consistent with secondary glaucoma; and (6) known history of glaucoma. Persons who were not included as having suspected glaucoma were classified as nonglaucomatous. 

Glaucoma cases were defined according to the International Society for Geographical and Epidemiologic Ophthalmology (ISGEO) criteria based on three categories. ^{21 –23} Category 1 cases were defined as optic disc abnormality (VCDR/VCDR asymmetry ≥97.5th percentile or neuroretinal rim width between 11 and 1 o'clock or 5 and 7 o'clock <0.1 VCDR) with a corresponding glaucomatous visual field defect. A glaucomatous visual field defect was defined as (1) glaucoma hemifield test (GHT) outside normal limits, and (2) a cluster of three or more nonedge, contiguous points, not crossing the horizontal meridian, with a probability of <5% of the age-matched normal on the pattern deviation plot on two separate occasions. Category 2 cases were defined as a severely damaged optic disc (VCDR or VCDR asymmetry ≥99.5th percentile) in the absence of an adequate visual field test. In diagnosing category 1 or 2 glaucoma, it was required that there be no other explanation for the VCDR finding (e.g., dysplastic disc or marked anisometropia) or visual field defect (e.g., branch retinal vein occlusion, macular degeneration or cerebrovascular disease). Category 3 cases were defined as subjects without visual field or optic disc data who were blind (corrected visual acuity, <3/60) and who had previous glaucoma surgery or had an IOP >99.5th percentile. 

Retinal Photography and Diabetic Retinopathy Assessment

Using a digital retinal camera (Canon CR-DGi with 10-D SLR back; Canon, Tokyo, Japan), two retinal photographs, centered at the optic disc and macula, were obtained from both eyes of each participant after pupil dilation, according to a standardized protocol. ^19,20 These photographs were sent to the University of Sydney and graded for retinopathy and other retinal diseases according to the Blue Mountain Eye Study protocol. ²⁴ Retinopathy was considered present if any characteristic lesion, as defined by the Early Treatment Diabetic Retinopathy Study (ETDRS) severity scale, was present: microaneurysms (MAs), hemorrhages, cotton wool spots, intraretinal microvascular abnormalities (IRMAs), hard exudates (HEs), venous beading, and new vessels. ²⁵ A retinopathy severity score was assigned for each eye according to a scale modified from the Airlie House classification system, described in detail elsewhere. ^19,20  

HRT II Imaging and Classification

HRT II testing was performed after pupil dilation in a dim room. After the test ended, the optic disc margin, defined as the inner edge of Elschnig's ring, was outlined by a trained ophthalmologist. Data were then analyzed with HRT software version 2.02. Images with a standard deviation higher than 50 μm were excluded. ²⁶  

The HRT II software automatically generates the linear discriminant functions (LDF1 described by Mikelberg et al., ²⁷ LDF2 described by Burk et al. ²⁸ ), and Moorfields regression analysis (MRA). ²⁹ The LDF1 is defined as F = (rim volume × 1.95) + (height variation contour × 30.12) – (corrected cup shape × 28.52) − 10.18, where the corrected cup shape = cup shape + (0.0019 × [50 − age]); the LDF2 is defined as F = 4.197 × (contour line height difference temporal − temporal superior) + (5.642 × contour line height difference temporal − temporal inferior) − (3.885 × temporal superior cup shape measure) − 0.974; the MRA is defined as rim area = 1.021 + 0.443 × disc area − 0.006 × age. The HRT II software compares the log of the measured neuroretinal rim area (globally and regionally) with the predicted rim area derived from a normative database, categorizing eyes as within normal limits (WNL, greater than or equal to 95% limit), borderline (BL, between 95% and 99.9% limit) or outside normal limits (ONL, below the 99.9% limit). The MRA is provided in six regions, as a global measure and as an overall MRA measure (equal to the worst global or regional category). In our study, the global MRA was a continuous measure (predicted minus actual global rim area), whereas the overall MRA was a binary measure that treated “borderlines” as positive results. We also calculated LDF3 described by Bathija et al., ³⁰ defined as F = (−4.37 × cup shape measure) − (5.57 × height variation along contour line) + (11.78 × mean RNFL thickness) + (1.85 × rim area) − 3.722803. The default cutoff points for the LDFs are the same, with a positive value indicating a normal disc and a negative value indicating a diseased disc. 

In addition, a support vector machine classifier with Gaussian kernel (SVM-Gauss) was used for our data because of its good performance in diagnosing glaucoma in another study. ³¹ Generally, a machine classifier is a computer algorithm that allows the computer to learn based on data and to make diagnostic decisions automatically. The SVM maps the multidimensional parameters into a high-dimensional feature space and creates a hyperplane to separate glaucomatous eyes from normal ones, maximizing the distance between all cases and the hyperplane. The global and six regional topographic measures for machine learning included disc area, cup area, rim area, cup/disc area ratio, rim/disc area ratio, cup volume, rim volume, mean cup depth, maximum cup depth, height variation contour, cup shape measure (CSM), mean RNFL thickness, RNFL cross-sectional area. In addition, global measures of horizontal cup/disc ratio, vertical cup/disc ratio, maximum contour elevation, maximum contour depression, contour line modulation (CLM) temporal-to-superior, and CLM temporal-to-inferior were included. 

Statistical Analysis

For persons with healthy eyes, one eye was randomly chosen, and for persons with DR, the eye with the more severe DR was chosen for analyses. Diabetes status and the presence of DR were analyzed as categorical variables as defined herein. HbA1c was analyzed as a continuous variable. Population characteristics were compared by using the Student's t-test or Mann-Whitney U test for continuous parameters and the χ² test for categorical ones, respectively. One-way analysis of variance (ANOVA) with Bonferroni adjustment was used for multiple comparisons (Stata, ver. 8.2; Stata Corp., College Station, TX, and R ver. 2.9.1, http://www.r-project.org/). 

For SVM classifier training, the whole data set was divided into a selection set and an evaluation set, with two thirds of the eyes being included in the selection set and one third in the evaluation set. This two-to-one split was repeated three times, and each time completely different original data were used as an evaluation set. Within the two-thirds selection data set, an internal 10-fold cross-validation was used, and the tuning method was applied for hyperparameter selection. 

A logistic regression model, described by Leisenring et al., ³² was used to assess the influence of diabetes or DR on the diagnostic precision of HRT algorithms. The HRT testing outcome (sensitivity or false-positive rate, dichotomized to 1 or 0) was treated as a dependent variable, whereas diabetes or its related factors were treated as independent variables. The overall MRA result (sensitivity or false-positive rate) does not require dichotomization as it is already a binary outcome. The other HRT II testing results derived from the global MRA, Mikelberg-LDF, Burk-LDF, Bathija-LDF, and SVM were continuous and needed to be dichotomized by fixing the specificity or false-negative rate. When we assessed the influence of diabetes/DR on the sensitivities for HRT algorithms, the specificities were fixed to a preset level (e.g., 80% or 90%) and then only data from glaucomatous patients were included in the logistic regression model. When we assessed the influence of diabetes/DR on the false-positive rates for HRT algorithms, the false-negative rates were fixed to a preset level (e.g., 40% or 60%) and then only data from persons without glaucoma were included in the logistic regression model. This statistical approach has the advantage of adjusting for many covariates and confounders (e.g., optic disc size and age), and was adopted by Medeiros and Zangwill and others, with further details described elsewhere. ^7,33 In our analyses, we constructed both the age-adjusted and multivariate logistic models (adjusted for age, sex, optic disc area, intraocular pressure, and axial length). To avoid potential confounding from misdiagnosed glaucoma, we excluded participants classified as having suspected glaucoma from the logistic regression analysis. Diabetic subjects who had macular edema (n = 15) were included in this study, given that our regression analyses showed that neither the RNFL thickness nor the sensitivity/false-positive rate for all HRT algorithms was associated with the presence of macular edema (all with P > 0.05). We did not note the presence of new vessels in the disc (NVD) in the eyes included in this study. 

Results

Of the 3280 subjects (response rate, 78.7%) who participated in the SiMES, 769 (224 without HRT II imaging data, 195 with HRT II testing but with poor image quality, 144 with retinal abnormalities other than DR, and 206 with a history of ocular surgery [including 154 with cataract surgery and 52 with evidence of previous retinal laser ablation seen from fundus images]) were excluded. Compared with those included (n = 2511), excluded participants (n = 769) were more likely to be older and to have a more negative spherical equivalent, cataract, visual impairment, diabetes, and DR (all with P < 0.05). Other characteristics were similar (e.g., the proportion with glaucoma was 4.9% among those excluded and 4.5% among those included). Among persons with glaucoma (n = 112), 86 (76.8%) were categorized as having no diabetes, 15 (13.4%) as having diabetes without DR, and 11 (9.8%) as having DR. Among persons without glaucoma (n = 2399), the number of patients in these categories was 1901 (79.3%), 358 (14.9%) and 140 (5.8%), respectively. Based on visual field testing with the Humphrey perimeter, the average mean deviation of the glaucomatous eyes was −7.47 dB (SD, 7.43; range, −30.95–1.01). 

We first assessed the relationships of diabetes and DR with HRT parameters by using both linear regression analysis and one way ANOVA. In multiple linear regression models after controlling for age, sex, IOP, axial length, and optic disc area size, the presence of DR was significantly associated with a thinner RNFL (β coefficient = −0.008; P < 0.001). The significant association persisted in all four quadrants (all with P < 0.01). Significant associations were also found between the presence of DR and increased CSM (data not shown). These findings were consistent with the data from one-way ANOVA with Bonferroni correction, illustrated in Figure 1. We also examined the relationship of HbA1c with RNFL thickness (by 12 quantiles) and CSM (by 12 quantiles). Although linear regression analyses showed no significant association of HbA1c with RNFL thickness and CSM (P > 0.05), there seemed to be a downward trend of RNFL thickness and an upward trend in CSM with increasing HbA1c after the 6th quantile (HbA1c = 5.9%). In addition, there appeared to be a downward trend in RNFL thickness when HbA1c was lower than 5.2% (1st quantile). Thus, the relationship between HbA1c and RNFL thickness appeared to be curvilinear. 

Figure 1.

 

Distribution of RNFL thickness (A) and CSM (B) by quadrant around the optic disc. Data are expressed as the mean. Error bar, 95% CI. (■) Persons without diabetes; (▩) persons with diabetes but without diabetic retinopathy; (□) persons with diabetic retinopathy. Tmp/Sup, temporal/superior quadrant; Tmp/Inf, temporal/inferior quadrant; Nsl/Sup, nasal/superior quadrant; Nsl/Inf, nasal/inferior quadrant.

We then described the areas under the receiver the operating characteristic (AUROC) curve for HRT II algorithms in this population. The AUROC was 73.4% (95% confidence interval [CI], 67.2–79.4) for the global MRA, 70.6% (95% CI, 65.9–75.2) for overall MRA, 74.2% (95% CI, 68.4–80.0) for Mikelberg-LDF, 71.4 (95% CI, 66.0–76.8) for Burk-LDF, 72.0% (95% CI, 66.2–77.9) for Bathija-LDF, and 82.1% (95% CI, 75.8–88.3) for SVM-Gauss. 

Finally, we assessed the influence of diabetes and DR on sensitivities and false-positive rates for the HRT II algorithms. In logistic regression models, the sensitivities for these HRT II algorithms were dichotomized as binary variables (by fixing the specificities at 80% or 90%) and then treated as dependent variables. Likewise, the false-positive rates for HRT II algorithms were dichotomized (by fixing the false-negative rates at 40% or 60%) and then treated as dependent variables. We found that (1) the presence of diabetes without DR was associated with neither the sensitivities nor the false-positive rates for the HRT II algorithms (all with P > 0.05, data not shown); (2) the presence of DR was not associated with the sensitivities for HRT II algorithms (Table 1); and (3) the presence of DR was significantly associated with the false-positive rates for Burk-LDF and Bathija-LDF (P < 0.05). The Burk-LDF and Bathija-LDF tended to yield higher false-positive rates for diagnosing glaucoma in persons with DR, compared with persons without DR. This difference was independent of optic disc size (Fig. 2). These associations persisted in different levels of fixed false-negative rate (40% and 60%; Table 2). The significant associations also persisted after adjustment for age, sex, optic disc size, IOP, and axial length. 

Table 1.

 

Logistic Regression Modeling of DR with the Sensitivities for HRT II Algorithms among Persons with Glaucoma

Presence of DR	Specificity Fixed at 84.0%		Specificity Fixed at 80%		Specificity Fixed at 90%
	Age-Adjusted OR (95% CI)	P †	Age-Adjusted OR (95% CI)	P †	Age-Adjusted OR (95% CI)	P †
Sensitivity for MRA*	0.89 (0.25–3.11)	0.86	0.70 (0.20–2.46)	0.58	0.63 (0.17–2.31)	0.49
Sensitivity for Mikelberg-LDF	0.58 (0.17–2.04)	0.40	0.77 (0.22–2.69)	0.68	0.97 (0.28–3.38)	0.96
Sensitivity for Burk-LDF	0.42 (0.10–1.75)	0.24	0.85 (0.24–3.05)	0.81	0.30 (0.06–1.49)	0.14
Sensitivity for Bathija-LDF	0.55 (0.15–2.06)	0.37	0.65 (0.18–2.33)	0.51	0.72 (0.19–2.71)	0.63
Sensitivity for SVM-Gauss	1.02 (0.98–1.07)	0.30	1.01 (0.96–1.06)	0.72	1.02 (0.97–1.06)	0.43

 OR, odds ratio; MRA, Moorfields regression analysis; Mikelberg-LDF, Mikelberg linear discriminant function; Burk-LDF, Burk linear discriminant function; Bathija-LDF, Bathija linear discriminant function; SVM-Gauss, support vector machine with Gaussian kernel.

*  When the specificity was fixed at 84.0%, the MRA was analyzed as an overall categorical variable (borderline and outside normal limits, 1; within normal limits, 0). When the specificity was fixed at 80% or 90%, the MRA was analyzed as a continuous variable (predicted minus actual global rim area).

†  Derived from age-adjusted logistic regression model. An OR of 1.0 indicates no influence of DR on the sensitivity; an OR <1.0 indicates a lower sensitivity than expected; and an OR >1.0 indicates a higher sensitivity than expected.

Figure 2.

 

False-positive rate of the Burk-LDF (A) and Bathija-LDF (B) tests, depending on optic disc area, stratified by the presence of diabetic retinopathy (DR). Data were derived from generalized additive models by fixing the false-negative rate to 42.9%.

Table 2.

 

Logistic Regression Modeling of DR with the False-Positive Rates for HRT II Algorithms among Persons without Glaucoma

Presence of DR	False-Negative Rate Fixed at 42.9%		False-Negative Rate Fixed at 40%		False-Negative Rate Fixed at 60%
	Age-Adjusted OR (95% CI)	P †	Age-Adjusted OR (95% CI)	P †	Age-Adjusted OR (95% CI)	P †
False-positive rate for MRA*	1.37 (0.88–2.12)	0.17	1.24 (0.79–1.94)	0.34	1.24 (0.79–1.95)	0.34
False-positive rate for Mikelberg-LDF	1.27 (0.81–2.00)	0.30	1.47 (0.81–2.68)	0.21	1.27 (0.83–1.96)	0.27
False-positive rate for Burk-LDF	1.41 (0.97–2.06)‡	0.07	1.61 (1.01–2.56)‡	0.03	1.60 (1.11–2.31)‡	0.01
False-positive rate for Bathija-LDF	1.71 (1.16–2.52)‡	0.007	1.93 (1.16–3.13)‡	0.007	1.63 (1.11–2.39)‡	0.01
False-positive rate for SVM-Gauss	1.49 (0.72–3.08)	0.29	0.98 (0.30–3.25)	0.98	1.28 (0.62–2.64)	0.50

 Abbreviations are as described in Table 1.

*  When the false negative rate was fixed at 42.9%, the MRA was analyzed as an overall categorical variable (borderline and outside normal limits, 1; within normal limits, 0). When the false-positive rate was fixed at 40% or 60%, the MRA was analyzed as a continuous variable (predicted minus actual global rim area).

†  Derived from the age-adjusted logistic regression model. An OR of 1.0 indicates no influence of DR on the false-positive rate; an OR <1.0 indicates a lower false-positive rate than expected; and an OR >1.0 indicates a higher false-positive rate than expected.

‡  Statistical significance (P < 0.05) remained in multivariate analyses after adjustment for age, sex, optic disc area, IOP, and axial length.

We performed several subsidiary analyses. After the exclusion of eyes with macular edema, the associations between the presence of DR with false-positive rates for Burk-LDF or for Bathija-LDF were similar (data not shown). We also performed statistical analyses by separately or collectively including the 154 pseudophakic eyes, 52 eyes with evidence of panretinal photocoagulation (PRP), and 76 eyes with suspected glaucoma. We found that a history of cataract surgery or PRP and the presence of suspected glaucoma were significantly associated with smaller rim area, thinner RNFL, and larger cup-to-disc area in multivariate analysis (all with P < 0.05). The presence of DR remained significantly associated with the false-positive rates for Burk-LDF and for Bathija-LDF (all with P < 0.05). 

Discussion

In this population-based study of adult persons aged 40 to 80 years, we report three main findings: First, diabetes had no effect on the sensitivity and/or specificity of HRT II tests. An increased HbA1c level may be associated with a decrease in RNFL thickness and with an increase in CSM, but this association was seen only for HbA1c levels higher than 6.0%. Second, persons with DR were more likely to have a thinner RNFL and larger CSM than were persons without diabetes. These associations were consistently observed in different quadrants around the optic disc. Finally, DR was significantly associated with the higher false-positive rates for Burk-LDF and Bathija-LDF. These findings indicate that while diabetes itself may not affect the performance of HRT II tests for diagnosing glaucoma, the presence of DR may be a source of false-positive HRT II test outcomes. 

Few studies have been conducted to examine the influence of diabetes on optic disc and RNFL measurements. A small case–control study based on HRT reported a thinner RNFL but an increased neuroretinal rim area in eyes with DR. ³⁴ Such an increase in neuroretinal rim area may be due to nerve swelling in eyes with advanced DR. ³⁵ In our study, however, most of the DR lesions were classified as minimal to moderate, and therefore we did not observe an increase in neuroretinal rim area. Our findings are also compatible with several clinic-based findings based on optical coherence tomography (OCT) or scanning laser polarimeter (GDx; Carl Zeiss Meditec). They found that RNFL thickness tended to decrease in persons with diabetes, even in diabetic patients without DR. ^{12 –14} In contrast, we did not find a significant reduction of RNFL thickness in diabetic patients without DR. Compared with patients obtained from hospital-based settings, diabetic patients without retinopathy identified from the population-setting are more likely to be in earlier disease stages and to have a shorter disease duration, and consequently, the retinal neurons may not have been affected. However, our study is likely to be more relevant to most individuals with diabetes in the community who are tested for glaucoma with the HRT II. 

The underlying mechanisms of retinal abnormalities in eyes with diabetes remain unclear. It has been shown that glial cells convey glucose to retinal neurons for the maintenance of its function, and thus excess glucose may be a contributing factor in neuronal abnormalities. ³⁶ This notion appears to be consistent with our findings of a decrease in RNFL thickness and an increase in CSM when the HbA1c was higher than 5.9%. Of note, it appeared that there was a decrease in RNFL thickness when the HbA1c was lower than 5.2% (1st quantile). This curvilinear relationship of HbA1c and RNFL thickness may warrant further confirmation. Advanced statistical methods (e.g., cosine transformation) may allow a more accurate delineation of the relation between HbA1c and RNFL thickness in future studies. 

To our knowledge, neither the influence of diabetes nor that of DR on the diagnostic ability of HRT II algorithms has been reported previously, largely because normal subjects in many clinic-based studies have strict inclusion criteria (without visual field defects and free of any eye disease, including cataract and DR). ^{27 –30} Sensitivity and specificity findings from these reports may represent ideal scenarios, but possible confounding effects from systemic or ocular conditions may be excluded, and thus these findings are unlikely to be generalizable to a true population-based setting. Given that diabetes has been identified as an important source of false-positive visual field findings, ¹¹ we chose not to use normal visual field results as a criterion for nonglaucomatous subjects. Excluding diabetic subjects with abnormal visual field results would probably result in exclusion of eyes with retinal structural abnormalities caused by diabetes. 

We note that not all the HRT II algorithms were affected by the presence of diabetes and DR. This effect can be explained on the grounds that different algorithms include different HRT parameters and that the aggregate effects of these parameters differ. Previous studies have shown that machine learning classifiers (e.g., SVM, neural network, and relevance vector machine) significantly outperform currently available algorithms for glaucoma detection. ^31,37 SVM is one of the classifiers that may provide improvement in glaucoma discrimination. ³⁸ This possibility is supported by our finding that the AUROC for SVM-Gauss (82.1%) was higher than the AUROCs (70%–74%) generated by other diagnostic algorithms. Our analyses also showed that the presence of diabetes or DR had no influence on the performance of SVM-Gauss (Tables 1, 2). These findings suggest that SVM can be used for glaucoma detection in patients with diabetes or DR. The presence of DR was not associated with the sensitivities of the HRT algorithms among persons with glaucoma (Table 1). These should be interpreted with caution, as our study included a relatively small number of glaucoma cases (n = 112, including 26 with diabetes). Such a sample size may not provide sufficient statistical power to detect the effect of diabetes among glaucoma cases. 

This study has several strengths, including a population-based sample and standardized grading of retinal photographs according to the Blue Mountains Eye Study protocol. Several limitations should also be considered. First, participants with poor HRT image quality and/or with signs of photocoagulation treatment for DR were excluded. Such exclusions would have reduced the sample size of people with DR and therefore would have reduced our statistical power. Second, DR was ascertained from only two digital images per eye so that the possibility of underdetection cannot be excluded. Such underestimation, nevertheless, is unlikely to be substantial as the DR prevalence in this population has been shown to be comparable or even higher in other studies in which retinal photographs of more than two fields were used to detect retinopathy. ^20,39 Finally, although our overall sample size was large, this study was limited by the relatively small sample size in each DR subgroup (by DR severity), and so we had limited statistical power to subdivide DR cases for supplementary analyses. It should be noted that some eyes with advanced DR could have had intraretinal fluid accumulation and exudates. These would lead to an increased neuroretinal rim area and RNFL thickness and so could have caused false-negative HRT test findings if these persons were affected by glaucoma. 

To conclude, we showed that diabetes per se does not affect HRT II algorithms for diagnosing glaucoma. However, the presence of DR is associated with false-positive rates for the Burk-LDF and Bathija-LDF algorithms, although not with false-positive rates for the MRA, Mikelberg-LDF, and SVM algorithms. These findings should be taken into account when HRT II is used to determine a structural diagnosis of glaucoma among persons with diabetes with signs of DR. 

Footnotes

 Supported by National Medical Research Council Grants 0796/2003 and Biomedical Research Council Grant 501/1/25-5.

Footnotes

 Disclosure: Y. Zheng, None; T.Y. Wong, None; C.Y.-L. Cheung, None; E. Lamoureux, None; P. Mitchell, None; M. He, None; T. Aung, None