| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
1From the Glaucoma Research Unit, Moorfields Eye Hospital, London, United Kingdom; and the 2Department of Optometry and Visual Science, City University, London, United Kingdom.
 |
Abstract |
|---|
 Top Abstract MethodsResultsDiscussionReferences |
|---|
METHODS. One hundred ninety-eight ocular hypertensive (OHT) and 21 control subjects were examined prospectively (1994–2001) with regular full-threshold Humphrey visual field (VF) testing. VF progression was assessed by point-wise linear regression (PLR) of sensitivity/time in Progressor for Windows software modified to include a novel spatial filter. Standard progression criteria (slope > –1 dB/year, P < 0.01) were applied to both "raw" (unfiltered) and "filtered" VF series. Three-omitting confirmatory VF criteria were also applied to unfiltered VF series. Specificity was estimated as the proportion of progressing control subjects and as the proportion of significantly improving subjects (both OHT and control) at the end of the study period.
RESULTS. Applying standard PLR, specificity was estimated at 91.8% to 97.5% using unfiltered standard PLR, compared with 93.5% to 98.4% using filtered standard PLR and 95.4% to 99.3% using unfiltered three-omitting PLR. The rate of identified VF progression in the OHT cohort was 32.3% with unfiltered standard PLR, 28.7% with filtered standard PLR, and 18.6% with unfiltered three-omitting PLR. There was no significant difference in time to detected progression between filtered and unfiltered standard PLR.
CONCLUSIONS. The use of confirmatory tests resulted in improved specificity using unfiltered data; however, application of the spatial filter resulted in similar specificity but with a higher rate of detected progression. This filter may therefore be useful in the monitoring of glaucomatous progression as it may reduce the dependence on confirmatory testing, although it has yet to be applied to longitudinal SITA data.
A potential strategy for reducing measurement variability, without recourse to additional testing or exclusion of "noisy" tests, is the post hoc application of a spatial filter to the VF data.10 11 Spatial filtering (or "processing") is adapted from digital image processing techniques. Medical images, such as those obtained by magnetic resonance imaging, may be digitally represented as a matrix of numerical values; measurement noise is reduced and image quality is improved by applying a mathematical process that exploits the spatial relationship between neighboring values. As the visual field may be considered as a similar numerical matrix, the same rationale may be applied to VF data. In this context, the measured threshold sensitivity of each test point within the VF is replaced by a "weighted" sensitivity value, which is estimated according to the magnitude of neighborhood sensitivity values. The first spatial filter applied to VF data, the Gaussian filter (based on a 3 x 3 test point grid), has been shown to reduce test–retest variability and to attenuate measurement noise.10 11 However, it has also been shown to attenuate useful signal, particularly small VF defects.12 This suggests that Gaussian filtering may be of limited benefit in the earliest stages of the disease process.
More recently, a novel spatial filter has been designed with the intention to incorporate the physiological relationship between measured sensitivity at all test points within the visual field.13 The filter was derived by examining the correlations and covariances between sensitivities among all pairs of VF test point locations, within 98,821 predominantly glaucomatous VFs.
The purpose of this study was to apply the novel spatial filter to real longitudinal VF data and to compare estimated specificity, progression rates, and time to progression, with and without the filter. The application of the filter was also compared to a VF progression technique that required multiple confirmatory testing (three-omitting point-wise linear regression analysis). The level of concordance in identified progression by VF and imaging techniques has been reported to be poor,14 15 16 and one explanation is that measurement noise results in failure to identify some progressing eyes and in the false detection of stable eyes. Noise reduction, therefore, may improve concordance. The effect of the filter on the concordance of VF progression with HRT structural progression was therefore also assessed.
|
Methods |
|---|
|
TopAbstractMethodsResultsDiscussionReferences |
|---|
All subjects in both groups had reliable and reproducible Humphrey full-threshold 24-2 VF test results (false positives and fixation losses <25% and false negatives <33%) at the time of recruitment. VF testing was performed every 4 to 12 months from the time of recruitment until September 2001. HRT imaging (Heidelberg Engineering, GmbH, Heidelberg, Germany) was performed on a yearly basis from 1994 to 1996, and subsequently at 4-month intervals until September 2001. In the present study, the same eye was selected for analysis as had been randomized in the original study. OHT eye randomization was stratified according "risk to glaucomatous conversion," using pattern electroretinogram, IOP and cup-to-disc area ratio at the time of recruitment,17 whereas control eyes were selected by simple randomization.
The study adhered to the tenets of the Declaration of Helsinki and had local ethics committee approval, as well as the subjects’ informed consent.
Visual Field Analysis
Each subject had a minimum of five Humphrey VF tests conducted over the study period, with the baseline test date taken as the date when the first HRT image was acquired. Five Humphrey VF tests were selected, as this is the minimum number of tests with which one may perform three-omitting point-wise linear regression (PLR; three tests to generate a linear regression, and two further confirmatory tests). To evaluate the performance of the spatial filter in a data set containing a range of VF variability, all available VF tests were included in the analysis irrespective of reliability criteria.
VF examinations from the subjects selected for analysis in this study were not used in the construction of the novel spatial filter.
VF progression was assessed by point-wise linear regression (PLR) of differential light sensitivity over time. This was performed using a version of Progressor (Institute of Ophthalmology, London, UK) which had been adapted to include the novel spatial filter. Each subject’s VF series was analyzed using the standard PLR criteria, the most commonly used PLR strategy in previously published studies.20 21 22 When using standard PLR criteria, a test point is identified as progressing if there is a slope of sensitivity over time exceeding –1 dB/year (P < 0.01). A test point is identified as improving if it satisfies the same criteria, except with a positive slope direction (+1 dB/year, P < 0.01). An eye was classified as progressing or improving when one or more test points satisfied the change criteria.
Standard PLR was performed on each VF series, first without application of the spatial filter (i.e., with raw VF sensitivity data, referred to as "unfiltered" VF data). The process was then repeated with the post hoc application of the spatial filter (referred to as "filtered" VF data).
The number of subjects identified as demonstrating significant progression or improvement at the end of the study period was compared by using each technique, as was the time to identification of significant change in these subjects. Ranges of specificity were estimated for both unfiltered and filtered VF data using two proxy measures—the number of control subjects demonstrating progression and the total number of subjects (both OHT and control) demonstrating improvement (VFs with one or more points changing with a slope of +1 dB/year at P < 0.01). All statistical analyses were performed with commercial software (Medcalc, ver. 7.4.2.0; Medcalc Software, Mariakerke, Belgium).
HRT Analysis
HRT mean topographies were generated and analyzed by using the Heidelberg Eye Explorer (ver. 1.7.0; Heidelberg Engineering GmbH, Heidelberg, Germany). Contour lines were drawn by a single observer (NGS) onto the baseline mean topographies and were exported automatically to the subsequent images. A manual alignment facility was used to correct contour line position if the automatically placed contour line was misplaced or if there was a magnification change.23 A minimum of five HRT mean topographies was available for each subject, with images of all qualities selected for analysis except where satisfactory contour line alignment could not be achieved. In total, eight mean topographies were excluded from the study, either as a result of double imaging or if the image was so grainy as to prevent adequate visualization of Elschnig’s ring.
The HRT progression strategy used in this study was based on linear regression analysis of sectoral rim area over time and has been described in detail elsewhere.15 Briefly, rim area (RA) was selected, as it has been shown to be a highly repeatable and reliable, clinically understandable stereometric parameter.2 3 All HRT mean topographies were analyzed by using the 320-µm reference plane, as this has been shown to result in less RA variability than the standard reference plane, which is the default reference plane used on the Explorer software platform.23 RA values for each Explorer-defined disc sector were calculated, and linear regression of these values over time was performed (recorded as percentage of baseline sectoral RA/time). Each sector series was classified as having high or low variability according to the residual SD (RSD) generated for each linear regression analysis, with the cutoff defined by the 50th percentile RSD value. To match the estimated specificity of both the unfiltered and filtered VF PLR strategies when comparing disc and field progression, low-stringency and high-stringency HRT progression strategies were adopted in this study. A low-stringency strategy has been defined in a previous study as a slope >1%/year (significance levels P < 0.05 for low-variability series and P < 0.01 for high-variability series); specificity was estimated at 88.1% to 90.5% for this technique, using the same subjects as in the present study.24 Specificity of 95.2% to 98.2% was estimated for the high-stringency strategy of significant slope >1%/year (significance levels P < 0.001 for low-variability series and P < 0.0001 for high-variability series).
Each OHT subject’s HRT series was examined using this technique, and the number of subjects showing significant change at the end of the study period was recorded. The number of subjects progressing by HRT was compared with the number of subjects progressing by VF, by using both unfiltered and filtered data.
The high-stringency HRT progression strategy has been compared with alternative PLR criteria—the three-omitting criterion—in a study performed on the same group of subjects as the present study.15 The three-omitting criterion is designed to increase specificity by including two confirmatory tests.24 A test point is identified as progressing if it satisfies standard PLR in each of three slopes. The first slope is constructed using all time points up to time point n, the second slope is constructed omitting point n and including the next point in the series (n + 1), and the final slope is constructed omitting points n and n + 1 with the inclusion of the next point in the series (n + 2). The use of the three-omitting criterion generated specificity estimates of 95.2% to 98.2% in the same two cohorts of subjects.15 The agreement between three-omitting PLR (unfiltered) and high-stringency HRT progression, in which the specificity estimates are closely matched at approximately 98%, was also compared in this study.
|
Results |
|---|
|
TopAbstractMethodsResultsDiscussionReferences |
|---|
|
|
Using the unfiltered VFs, 64 subjects with OHT (positive detection rate 32%, 95% confidence interval 26%–39%) were identified as having progressive disease, of whom 24 also showed progression by the HRT low-stringency strategy. A further 42 (21%) subjects progressed by HRT alone (Fig. 1) . Fifty-seven subjects with OHT (positive detection rate 29%, 95% confidence interval 23%–36%) showed progression by filtered VFs, and seven also showed progression by the HRT high-stringency strategy. In a further 17 (9%) subjects, disease progression was shown by HRT alone (Fig. 2) .
|
|
|
|
|
|
|
|
Discussion |
|---|
|
TopAbstractMethodsResultsDiscussionReferences |
|---|
Confirmatory tests are an approach to counteract the intertest measurement variability, or long-term fluctuation, which represents the primary barrier to the identification of true change within a longitudinal series of VFs.25 26 Additional tests to confirm progression have the benefit of lessening the effect of poor overall performance in a VF test in a series (which may affect many points in the VF), but with the caveat that increased costs are incurred due to additional visits.27 It is likely that confirmation tests and spatial filtering will prove to be complementary. The concept of spatial processing represents an attractive proposition because it does not require the collection of further test data or any modification to the testing process itself, as it is applied post hoc to previously acquired VF data. The first spatial filter applied to VF data—the Gaussian, or simple averaging, filter—has been shown to be capable of dampening the effect of long-term variability. Its performance is unsatisfactory, however, where localized but significant VF loss exists. In this circumstance, the field loss may be obscured by the spatial processing technique. Spry et al.12 observed that Gaussian filtering, when applied to simulated VF data resulted in a modest specificity gain but considerable sensitivity depreciation for small, progressive VF losses; a small specificity loss was also observed for large, progressive defects. In their study, temporal processing, effectively a running average of threshold sensitivity over time, was propounded as a more predictable method of increasing sensitivity gain. This technique exerts a "smoothing" effect of variability over time. Its benefit is decreased with an increasing number of available tests. Unlike the spatial filter used in the present study, neither the temporal nor the Gaussian filter was designed along physiological principles.
There was likely a small improvement in specificity when the novel spatial filter was applied, and this was associated with a similar decline in the proportion of eyes identified as progressing. Gardiner et al.13 originally tested the novel spatial filter by using VF computer simulations that were based on robust and realistic estimates of the visual field noise.4 28 Localized defects for each point in the visual field were tested, including some consisting of just two progressing points. Results indicated that, as expected, the Gaussian filter blurred many of the progressive defects, whereas the new filter improved detection rate in >90% of the defects tested.
Specificity estimates approaching 100% have been achieved previously for PLR techniques, without spatial filtering, using glaucomatous VF data with simulated measurement variability.29 The progression criteria applied were far more rigid than those used in the present study, requiring a regression slope of –1.0 dB/year, with a significance at P < 0.01 in the same three test points in three of four consecutive tests. By excluding the necessity for "cluster" test point progression, the PLR technique used in the present study may be predisposed to lower specificity because of the wide range of variability at individual test locations. Lower specificity may also be expected, as no confirmatory tests were included to counter long-term fluctuation. In the absence of an independent gold standard based on which to classify a subject as having progressed, estimates of specificity were derived from surrogate measures based on two assumptions. First, that threshold sensitivity should not improve over time and, second, that threshold sensitivity decay over time should not exceed age-related decay in control subjects. With respect to the former assumption, all subjects included in the analysis had reproducibly normal and reliable VFs at baseline and were therefore unlikely to exhibit prolonged learning effects over time, which may have resulted in some positive threshold sensitivity change.17 With respect to the latter assumption, it is possible that, as individuals may age at different rates, some threshold sensitivity loss due to normal aging may have been flagged as progression in the control cohort. However, no progression was seen in the normal subjects after the application of the novel spatial filter.
Likewise, in the absence of an independent gold standard for disease progression, it is not possible to obtain a direct measure of test sensitivity. Given the high estimates of specificity, both for unfiltered and filtered PLR, it is very likely that the great majority of "progressors" identified by either technique are likely to represent true disease progression. Spatial processing resulted in a small decrease in detected progression or positive hit rate of 3.6%, which may equate to a relatively small diminution of sensitivity. In the present study, progressing defects were expected to be smaller, as the subjects were ocular hypertensive with normal VF test results at baseline. In a previous study with the Gaussian filter, a sensitivity loss of up to 50% was estimated for small progressive defects (two progressing points) at a true progression rate of –1 dB/year using 10 visual field tests, compared with up to 20% for larger defects (18 progressing points).12
A particular problem identified in the VF progression techniques adopted in large-scale clinical trials is the number of false-positive results. Reversal of progression was examined in subjects from the Ocular Hypertension Treatment Study (OHTS). Only 12% of VF test results returned to normal after three consistent abnormal VF test results, compared with 66% when only two abnormal VF results were used.30 This finding suggests that, for the VF progression criteria used in OHTS, the adoption of two additional confirmatory VF tests would improve stability and specificity. A high false-positive detection of field progression is also suggested when a single test point is used to flag progression by PLR, as in the present study.22 31 The high specificity achieved by application of the novel spatial processing technique suggests that the number of false-positive progressors would be limited. Whether this removes the necessity for confirmatory testing is not clear. Within the original betaxolol versus placebo study period (1994–1998), a change in AGIS VF score from 0 to greater than 1 at the same test point location on three occasions was necessary to confirm progression.17 Individuals who were suspected of disease progression were therefore subjected to episodes of increased frequency of VF testing, during which the investigators were seeking to confirm progression. In general, subjects were tested at a frequency of three times per year, which has previously been identified as an appropriate frequency by which to detect progression.24 A recent study has suggested, however, that an adaptive test interval, shortened at periods when progression is suspected, may detect progression earlier than at a fixed interval rate.32 An indicator that application of the spatial filter may limit the requirement for confirmatory testing is the observation that it achieves similar estimates of specificity as the three-omitting technique.15 However, the use of confirmatory criteria is at the cost of a significant decrease in the positive detection rate compared with filtered standard PLR.
The poor level of agreement with structural progression is consistent with previous observations in subjects with glaucoma,14 in those with OHT,15 and with the use of alternative ONH imaging technologies such as optical coherence tomography.16 Although there is a widely held view that structural changes are detectable before functional changes in progressive glaucoma,33 it may be a manifestation of the limitations of the detection techniques available. Explanations for poor structural and functional correlation may include structural changes occurring without concomitant functional change (such as lamina cribrosa bowing) and functional changes occurring without structural alteration (such as ganglion cell dysfunction). Another possible theory is that the poor correlation arises from differences in the amount of measurement variability between the two testing modalities. It is, however, difficult to identify improvement in agreement using the spatial filter because of the large reduction in hit rate using the HRT with the more stringent progression criteria. In comparing the filtered VF results with those obtained with the low-stringency HRT strategy, albeit with the caveat that the specificities are not matched, one can observe that there is very little reduction (and no improvement) in the agreement between disc and field progression with the application of the filter.
The novel spatial filtering technique used in this study, which was designed to mimic the physiological relationship between test point pairings within the VF, has been shown to achieve similar specificity as using PLR with confirmatory criteria, with an increased rate of detected progression. The filter, therefore, may be useful in the detection of glaucomatous progression and may be suitable in examining data from clinical trials. As it is applied post hoc to data that have already been collected, it does not require additional cost in terms of clinic time and repeat VF testing. However, the filter was constructed and tested using full-threshold VF tests (albeit from a different group of subjects). It has yet to be applied to VF series acquired using the Swedish Interactive Thresholding Algorithm (SITA). It may be necessary for the spatial filter to be reconstructed using a database of SITA fields to obtain optimal performance in that context.
|
Footnotes |
|---|
Submitted for publication May 28, 2006; revised August 20, 2006; accepted November 21, 2006.
Disclosure: N.G. Strouthidis, Heidelberg Engineering (F); A. Scott, None; A.C. Viswanathan, None; D.P. Crabb, None; D.F. Garway-Heath, Heidelberg Engineering (F), Carl Zeiss Meditec (C, F)
The publication costs of this article were defrayed in part by page charge payment. This article must therefore be marked "advertisement" in accordance with 18 U.S.C. 
1734 solely to indicate this fact.
Corresponding author: David F. Garway-Heath, Glaucoma Research Unit, Moorfields Eye Hospital, 162 City Road, London, EC1V 2PD, UK; david.garway-heath{at}moorfields.nhs.uk.
|
References |
|---|
|
TopAbstractMethodsResultsDiscussionReferences |
|---|
This article has been cited by other articles:
|
|
 |
|
  A. Kotecha, A. Spratt, C. Bunce, D. F. Garway-Heath, P. T. Khaw, A. Viswanathan, and the MoreFlow Study Group Optic Disc and Visual Field Changes after Trabeculectomy Invest. Ophthalmol. Vis. Sci., October 1, 2009; 50(10): 4693 - 4699. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
