AIM To evaluate the ability of structural parameters (as determined by retinal nerve fibre layer (RNFL) measurements obtained with the scanning laser polarimeter (SLP-NFA/GDx)) and functional parameters (as determined by automated perimetry) to discriminate between normal and glaucomatous eyes.
METHODS In a case-control study, a total of 91 normal subjects and 94 patients with glaucoma underwent automated perimetry and RNFL measurements obtained with the SLP. Three independent scans of each eye were obtained and a mean image was created and used for further analysis. Only one eye per individual was randomly included in the study. The sensitivity (Se) and specificity (Sp) of 12 RNFL parameters were calculated according to the SLP internal normative database. The Se and Sp of the visual field (VF) global indices and the glaucoma hemifield test (GHT) were also calculated according to the instrument's normative database. Receiver operator characteristic (ROC) curves were built for each SLP parameter and VF index. Fisher's linear discriminant formulas (LDFs) were developed for VF indices (VF LDF), SLP measurements (SLP LDF), and both examinations (combined LDF).
RESULTS According to the SLP internal database, the parameters with better Se and Sp were: superior/nasal ratio (Se = 58.5%; Sp = 86.8%), and GDx the number (Se = 43.3%; Sp = 96.7%). The construction of an ROC curve for the number resulted in Se = 84% and Sp = 79%. The creation of LDFs improved both the sensitivities and specificities when compared with isolated parameters SLP LDF (Se = 90.4%; Sp = 82.4%), VF LDF (Se = 89.4%; Sp = 89.0%), and combined LDF (Se = 93.0%; Sp = 90.1%). The sensitivity to diagnose early and moderate glaucomatous damage observed with the GHT was lower than that obtained with the number (p<0.01).
CONCLUSIONS Creation of LDFs enhanced the Se and Sp for both VF and SLP. Integration of SLP and VF in a combined LDF reached the highest Se/Sp relation, suggesting that these examinations may be additive concerning the diagnosis of glaucoma. The SLP parameter the number may be more sensitive than the GHT in diagnosing early and moderate glaucomatous damage.
- scanning laser polarimetry
- visual field
Statistics from Altmetric.com
There has been increasing evidence that retinal nerve fibre layer (RNFL) analysis may improve the diagnosis of early glaucomatous damage.1-4 Hoyt and Newman5 first suggested that localised RNFL defects represent the earliest detectable defect in glaucoma. Sommer et al6demonstrated that RNFL abnormalities may be detected up to 6 years before the development of visual field loss. Other authors have stressed that RNFL examination is a useful way of monitoring glaucomatous damage.78
Scanning laser polarimetry (SLP) measurement is based on the birefringent properties of the RNFL, which has its microtubules disposed in an organised, parallel fashion. This peculiar anatomy leads to a change in the state of polarised light as it passes through the RNFL, creating a retardation that is directly proportional to its thickness.9 Experimental and clinical studies have demonstrated that SLP provides quantitative910 and reproducible11 measurements of the RNFL thickness, shown to be significantly lower among glaucomatous10 and ocular hypertensive1213 eyes when compared with healthy eyes, although considerable overlap is present between these groups.
Most of the studies investigating glaucoma require both anatomical and functional damage to define the disease, although a time interval exists between these.1 Hence, we designed a study in order to compare the ability of SLP structural parameters and automated perimetry functional measurements (used separately or in conjunction) to discriminate between normal and glaucomatous eyes.
After approval of the ethics committee of the University of Campinas and written consent, all subjects underwent a thorough ophthalmological examination including slit lamp biomicroscopy, applanation tonometry (Goldmann), gonioscopy, dilated retinal and optic nerve head examination (20 dioptre and 78 dioptre lenses), automated perimetry using the Humphrey field analyser II, program 24-2, full threshold strategy (Humphrey Systems, Dublin, CA, USA) and SLP, using the nerve fibre analyser-GDx (Laser Diagnostic Technologies, San Diego, CA, USA), software version 1.0.12.
The inclusion criteria for both the normal and glaucoma groups were best corrected visual acuity ⩾20/30, spheric equivalent ⩽ + or −5 dioptres, age over 40 years (no upper limit), and two reliable Humphrey 24-2 visual fields.14 Patients in both groups were excluded if presented with history of systemic or ocular disease (except glaucoma) that could interfere with the scanning laser polarimeter or visual field results. We also excluded pseudophakic eyes and those with significant cataract—greater than mild lens opacification, according to the Lens Opacity Classification System III.15 One eye per patient was randomly selected if both eyes were eligible.
Normal subjects were recruited from volunteers among the medical staff, university members, spouses, and friends of patients. These were excluded if presented with pressure above 22 mm Hg or with a suspicious disc (that is, localised rim loss, optic disc haemorrhage, cup/disc asymmetry >0.2), or with glaucomatous visual field defects (as defined below).
Glaucoma patients were recruited from the glaucoma service of the University of Campinas. The inclusion criteria were: clinical diagnosis of primary open angle glaucoma with two or more IOP measurements above 22 mm Hg, optic nerve examination demonstrating glaucomatous damage (that is, cupping, localised rim loss, disc haemorrhage, or cup/disc asymmetry > 0.3). Furthermore, a typical glaucomatous visual field defect had to be present on at least two reliable VF testings. This was defined as: (a) two or more contiguous points with a 10 dB loss or greater in the superior or inferior Bjerrum areas, compared with perimeter defined age matched controls; (b) three or more contiguous points, with a 5 dB or greater loss, in the superior or inferior Bjerrum areas.14 Neither the global indices nor the GHT were used to define glaucomatous or normal visual fields.
Patients were classified as having early, moderate, or severe glaucomatous damage according to the following criteria16: (a) early damage: MD no worse than −6 dB and CPSD no worse than 1%; (b) moderate damage: MD between −6 dB and −15 dB, and CPSD no worse than 1%; and (c) severe damage: MD worse than −15 dB or CPSD worse than 1%.
To obtain RNFL measurements of each eye, we created a mean image, which was the average of three consecutive 15° images, centred on the optic nerve. Room lights were kept on and pupils were left undilated. In order to be included, each single image had to pass the software's criteria of quality. A single examiner (RLP) obtained the images and delineated the disc margin.
The following RNFL parameters were collected: symmetry, superior ratio, inferior ratio, maximum modulation, average thickness, superior to nasal ratio, ellipse modulation, ellipse average, superior average, inferior average, superior integral, and the number. The meaning of each RNFL parameter has been reported elsewhere.16
The sensitivity and specificity of each single parameter were calculated according to the GDx normative database. Values outside the 95% confidence limits of the internal software database, automatically labelled as “outside normal limits”, were considered abnormal. The neural network variable, the number, was considered abnormal if greater than 70 (as suggested by the manufacturer). Additionally, an “all normal” variable was created and defined as abnormal if at least one of the 12 parameters measured by the NFA was “outside normal limits”.
Likewise, the sensitivity and specificity of the visual field (VF) global indices (MD, SF, and CPSD) and the glaucoma hemifield test (GHT) were calculated. The GHT was considered abnormal when the printout displayed the message “outside normal limits”, which refers to a value falling outside the 99% confidence interval. Additionally, MD, SF, and CPSD were considered abnormal when labelled with p <5% according to the Humphrey's database (Humphrey field analyser's II statistical software, statpac 2).
Subsequently, receiver operator characteristic (ROC) curves were plotted for each individual SLP parameter and VF index. The area under the curve was calculated and a z test was employed to determine if the area was significantly different from 0.5 (which indicates no discrimination).17 Cut-off points were arbitrarily selected to determine the best sensitivity/specificity relation for each single VF and SLP measurement.
Fisher's linear discriminant functions (LDFs) composed of the best combination of SLP and VF measurements were created. The ability to discriminate between normal and glaucomatous eyes with the SLP LDF, the VF LDF, and the combined (VF+SLP) LDF was tested.
Comparisons were made with the Student's t test (for continuous variables), the χ2 test (Yates's corrected), and the Fisher exact test (for the categorical data). p Values of less than 0.05 were considered significant.
A total of 185 eyes, 91 normal and 94 glaucomatous, were enrolled in this study. Among the 94 glaucomatous eyes, 14 were classified as having early glaucomatous damage, 30 had moderate, and 50 had severe damage (mean MD = −11.09 (SD 9.10) dB). Demographic data are displayed in Table 1. The mean age in the glaucoma group (64.7 (11.3)) was significantly higher than the normal group (47.3 (6.0)) (p<0.05). The cup/disc ratio was significantly higher among glaucoma patients (0.69 (0.12)) when compared with controls (0.32 (0.12)) (p<0.05). Because of the significant differences in age and cup/disc ratio, these were included as candidate variables in the linear discriminant function (LDF) analyses. The stepwise analysis, used to create the LDFs, demonstrated that both age and cup/disc ratios were weak variables, and therefore were not included in the final structural LDF formula.
There were statistically significant differences (p<0.001) between healthy and glaucomatous eyes for all VF global indices and SLP parameters, except for symmetry (p = 0.12) (Tables 2 and 3) (Fig 1). When the NFA/GDx normative database was used, SLP parameters showed low sensitivities (Se) and high specificities (Sp) for the diagnosis of glaucoma (Table 4). Conversely, the “all normal” variable resulted in a sensitivity of 79.8% and a specificity of 56%. Based on the Humphrey's internal database, VF global indices showed moderate sensitivities (over 70%), except for SF (Se = 33.0%), and high specificities (above 90%).
Areas under the ROC curves for all SLP parameters and VF indices were highly significant (p< 0.001), except for symmetry (p = 0.184) (Table5, Fig 2). Selection of new cut-off points improved the sensitivity/specificity relation of all SLP parameters when compared to the results obtained with the GDx normative database. Individual parameters found to have better sensitivity/specificity relation were the number (Se = 84%; Sp = 79%), ellipse modulation (Se = 82%; Sp = 73%), and superior/nasal ratio (Se = 82%; Sp = 79%). Although highly significant ROC curves were observed for each VF parameter (p<0.001), we could not find, in our model, cut-off points to improve the sensitivity/specificity relation for SF and CPSD (Table 5).
A statistically significant area of 0.961 under the ROC curve was obtained for the VF LDF (p<0.01). With a cut-off set at lower than 0.430 as abnormal, a sensitivity of 89.4% and a specificity of 89.0% was observed, with a labelling accuracy of 90.3% on our population sample. The VF LDF formula derived from the best set of parameters was:
VF LDF = −1.679 − (0.072 × MD) + (0.217 × SF) + (0.211 × CPSD)
Likewise, the best formula achieved with SLP parameters was:
SLP LDF = − 3.131 + (0.994 × ellipse modulation) − (0.017 × the number) − (0.086 × average thickness) + (0.111 × ellipse average)
The ROC curve for this formula resulted in a statistically significant area of 0.930 (p <0.01). With a cut-off point set at lower than −0.547 as abnormal, it showed a sensitivity of 90.4%, and a specificity of 82.4%, with an accuracy of 85.9%.
The use of a combined discriminant function resulted in an increase of the sensitivity/specificity relation. When the cut-off point was set at higher than −0.446 as abnormal, it showed an area under the ROC curve of 0.98, a sensitivity of 93.0%, a specificity of 90.1%, and an accuracy of 90.8%.
Combined LDF = 1.015 − (0.645 × ellipse modulation) + (0.017 × the number) − (0.24 × ellipse average) + (0.207 × SF) + (0.196 × CPSD)
Table 6 discloses the different sensitivities obtained with the GHT, the number (cut-off = 32), the SLP, and VF LDFs according to the severity of glaucomatous damage. The sensitivities obtained with the number and the structural LDF for the diagnosis of early/moderate glaucomatous damage were significantly higher than that obtained with the GHT (p = 0.009, and p = 0.0001, respectively).
In the present study, RNFL retardation measurements obtained with the SLP were significantly higher in normal eyes compared with glaucomatous eyes, although there is a considerable overlap between both groups. The symmetry parameter was the only measurement not significantly different between the groups (p = 0.12). These findings are consistent with previous studies that stressed the large range of normal mean values of retardation.101618-20
Although the mean age was different among the groups, it did not influence our LDF analysis to discriminate between normal and glaucomatous eyes, as it proved to be a weak variable to be used in the LDF formula.
All individual SLP parameters showed low sensitivities and high specificities when the NFA/GDx database was used as reference and none of the individual RNFL measurements could correctly assign as outside normal limits the majority of glaucoma cases. There are two possible explanations for this finding. Firstly, the normative database may be insufficient (in number or in variety), which may be influenced by the variability of the axis of corneal polarisation as recently reported by Greenfield et al.21 The SLP employs an anterior segment compensating algorithm, which assumes a fixed axis of corneal polarisation, and may not be adjusted to such physiological variability. Secondly, the use of the 95% confidence interval to define “outside normal limits” may be too stringent, decreasing the sensitivity in spite of a high specificity. As it is, the database may be useful to confirm the diagnosis of glaucoma, but not to detect early glaucomatous damage.
Construction of ROC curves for individual structural parameters and arbitrary selection of cut-off points significantly enhanced the sensitivity of all SLP parameters, reaching values around 80% for the number, ellipse modulation, superior/nasal, and maximum modulation. It is possible that modulation parameters and ratios are not greatly sensitive to changes in the axis of corneal polarisation, explaining the better sensitivity/specificity relation obtained with this parameter.22 The number was the SLP parameter with the largest area under the ROC curve (0.87), leading to a sensitivity of 84%, a specificity of 79%, and an accuracy of 80.5%, with a cut-off set at 32. When analysing the same variable, Weinrebet al16 were able to build a ROC curve with an area of 0.78 when the cut-off point was set at 27, which resulted in a specificity of 82% and a sensitivity of 62%. The use of a cut-off point that would increase the sensitivity is desired in a test designed to provide early diagnosis of glaucoma, yet lower cut-off points for the number lead still to only moderate sensitivities in both studies. Similarly, other authors found that a cut-off for the number at around 30 performed better than at the suggested 70.23
On the other hand, the analysis of functional parameters showed that, except for SF, VF indices showed better sensitivities than SLP measurements, according to the respective equipment databases. MD and CPSD were found to have sensitivities of 83% and 80%, respectively. The reason for the better sensitivities obtained with VF indices may be secondary to the large Humphrey's statistical database, or to a selection bias in our criteria for defining glaucoma (although both structural and functional damage were required for the patient to be included). The use of clusters of low sensitivity points may have artificially increased the sensitivity and specificity of indices such as the CPSD and the GHT, designed to detect focal differences or regional asymmetries. Nevertheless, the 72.3% sensitivity of the GHT shown in this series is similar to the 78% sensitivity reported by Asman and Heijl,24 and lower than that previously reported by Katz et al (Se = 94%).25However, the latter study defined typical glaucomatous defects based on Goldmann's manual perimetry, which may have led to the inclusion of more advanced stages of glaucoma, artificially increasing the sensitivity
The GHT has been described by some authors as the best single objective statistical criterion for the detection of glaucoma.2627The 72.3% sensitivity observed with the GHT was lower than that obtained with the number when its cut-off point was set at 32 (Se = 84.0%). This difference becomes greater when the sensitivities of both parameters are compared in the early/moderate damage group (the number = 75%; GHT = 45.4%) (p = 0.005). The SLP LDF (Se = 86.3%) was also more sensitive in diagnosing early and moderate glaucoma when compared to the GHT (p<0.001) (Table 6). These findings indicate that the SLP neural network may be capable of showing change earlier than classic functional measures, consistently with the description that RNFL defects may occur long before visual field damage.56
In this study, the creation of a VF LDF (Se = 89.4%; Sp = 89.0%) and a SLP LDF (Se = 90.4%; Sp = 82.4%) resulted in similarly high sensitivities and specificities, higher than any individual SLP parameter or VF index. When applied to the early and moderate subgroup, the SLP LDF reached slightly higher sensitivities than VF LDF, although the differences were not significant (p = 0.395).
The use of functional and structural variables in a combined LDF improved the overall sensitivity/specificity relation, reaching a 92.5% sensitivity and a 90.1% specificity. Enhancement of sensitivity and specificity in this model suggests that VF and SLP parameters may be additive in the diagnosis of glaucoma. These findings are consistent with those reported by Caprioli,14 who described that the combination of functional (VF indices) and structural parameters (computer derived variables of the optic nerve and RNFL indirect measures) in a linear discriminant function improved the capability to diagnose glaucoma. This method of analysis correctly classified 87% of the studied eyes, compared to a 77% precision with a functional LDF, and a 76% precision of a structural LDF. Higher labelling accuracies were achieved with the LDFs developed in the present study (90.3% for the functional LDF, 85.9% for the structural LDF and 90,8% for the combination of structural and functional parameters).
Although the LDFs were not tested in a independent sample, we estimated the bias of the ROC curve with a k-fold cross validation on randomly chosen training variables and the test sets of variables. This approach resulted in an estimated decrease of 1.6% in the area under the structural LDF ROC curve and of 1.8% for the combined LDF.
Weinreb et al16 also developed a structural LDF, based on SLP readings that could increase the ability to differentiate between normal and glaucomatous eyes with early to moderate visual field damage (ROC curve of 0.89, a sensitivity of 74% and a specificity of 92%). The application of the LDF proposed by those authors in our sample resulted in an overall sensitivity and specificity of 85.1% and 83.5%, respectively.
Furthermore, that formula resulted in a sensitivity of 77.2%, when used in our early and moderate glaucoma groups (defined with the same criteria as Weinreb et al). These results are similar to those reported in that original series (Se = 74%) and apparently validate that formula. Nevertheless, the estimated sensitivity of our SLP LDF (86.3%) was higher than Weinreb's when applied to the same subgroup of patients, although the difference was not statistically significant (p = 0.269). The lack of significance may be due to the small number of early/moderate glaucoma patients in the present study. Further evaluations on larger samples are needed to address this issue. Interestingly, the two formulas share three variables in common: average thickness, ellipse modulation, and ellipse average. The fourth variable included in our LDF, the number, took part in only one of the best set of 20 models of three predictors used in Weinreb's LDF.
Direct comparisons between the present study and previous reports of SLP performance to detect glaucoma may not be straightforward, in part because population samples seem to have distinct spectra of glaucomatous damage. The 96% sensitivity and 93% specificity reported by Tjon-Fo-Sang and Lemij28 when using the SLP to detect glaucoma were obtained with a previous version of the NFA. The calculation of mean retardation was processed in six selected areas and avoided areas crossed by vessels. Although their study included patients with a range of MD loss (mean MD = −10.33 dB, ranging from −31.5 to 0.76 dB) similar to ours (mean MD = −11.09 dB, ranging from −27.6 to 4.16 dB), the GHT was necessarily outside normal limits, which was not an inclusion criterion in our study. The use of the NFA-GDx instead of the original NFA may account for some differences in the results, particularly because the normative database and the methods employed by Tjon-Fo-Sang and Lemij are different from the data analysis included in the GDx software used in this study. Additionally, new features of the NFA-GDx were developed in order to evaluate the quality of the captured image and to reduce interoperator and intraoperator variability of the image acquisition.
In conclusion, this study showed that integration of multiple SLP variables on a LDF improved the ability to differentiate between normal and glaucomatous eyes more than any individual SLP parameter. The application of a linear discriminant model also suggested that visual field indices and SLP parameters may be additive concerning the diagnosis of glaucoma. Further studies are necessary to validate these LDFs in different population samples. Finally, as far as we are aware, this is the first report to indicate that the SLP parameter the number may be more sensitive to the detection of glaucomatous damage than the GHT, especially in the early/moderate damage group.
The authors have no proprietary interest in any product discussed in this presentation or a competing instrument.
Presented in part at the Association for Research in Vision and Ophthalmology (ARVO) annual meeting 2000.
Supported by a grant from Fundação de Amparo a Pesquisa do Estado de São Paulo FAPESP (No 9698692), São Paulo, Brazil.