Short-term Detection of Fast Progressors in Glaucoma: The Fast-PACE (Progression Assessment through Clustered Evaluation) Study

Abstract

Purpose:

To evaluate the performance of an intensive, clustered testing approach in identifying eyes with rapid glaucoma progression over 6 months in the Fast-PACE (Progression Assessment through Clustered Evaluation) study.

Design:

Prospective cohort study.

Participants:

125 eyes from 65 primary open-angle glaucoma (POAG) subjects.

Methods:

Subjects underwent two sets of 5 weekly visits (clusters) separated by an average of 6 months, then were followed with single visits every 6 months for an overall mean follow-up of 25 months (mean of 17 tests). Each visit consisted of testing with standard automated perimetry (SAP) 24–2 and 10–2, and spectral-domain optical coherence tomography (SD OCT). Progression was assessed using trend analyses of SAP mean deviation (MD) and retinal nerve fiber layer (RNFL) thickness. Generalized estimating equations were applied to adjust for correlations between eyes for confidence interval (CI) estimation and hypothesis testing.

Main Outcome Measures:

Diagnostic accuracy of the 6-month clustering period to identify progression detected during the overall follow-up.

Results:

19 of 125 eyes (15%, CI: 9%−24%) progressed based on SAP 24-2 MD over the 6-month clustering period. 14 eyes (11%, CI: 6%−20%) progressed on SAP 10–2 MD, and 16 (13%, CI:8%−21%) by RNFL thickness, with 30 of 125 eyes (24%, CI:16%−34%) progressing by function, structure, or both. Of the 35 eyes progressing during the overall follow-up, 25 had progressed during the 6-month clustering period, for a sensitivity of 71% (CI: 53%−85%). Of the 90 eyes that did not progress during the overall follow-up, 85 also did not progress during the 6-month period, for a specificity of 94% (CI: 88% – 98%). Of the 14 eyes considered fast progressors by SAP 24–2, 10–2 or SD OCT during the overall follow-up, 13 were identified as progressing during the 6-month cluster period, for a sensitivity of 93% (CI: 66% – 100%) for identifying fast progression with a specificity of 85% (CI: 77% – 90%).

Conclusion:

Clustered testing in the Fast-PACE study detected fast-progressing glaucoma eyes over six months. The methodology could be applied in clinical trials investigating interventions to slow glaucoma progression and may also be of value for short-term assessment of high-risk subjects.

PRÉCIS

In a prospective longitudinal study, an intensive, clustered testing approach was able to detect a substantial proportion of eyes with rapid glaucoma progression within a short follow-up period of 6 months.

Glaucoma is the leading cause of irreversible blindness in the world.¹ Disease severity at diagnosis and rate of progression are two critical factors in determining who gets visually impaired or blind from glaucoma.² For most patients, the disease is considered to have a protracted course, generally taking many years to display significant deterioration.^3–5 That said, some progress rapidly, and these individuals are at high risk for disability, underscoring the need for prompt detection and treatment of rapid progression.

Reduction of intraocular pressure (IOP) has been shown to slow down disease progression, but some patients continue to progress despite well-controlled IOP.^6,7 Neuroprotective therapies aimed at directly protecting neural tissue from damage via non-IOP mechanisms would likely be essential to prevent blindness in such cases. However, the development of such therapies has been hindered by difficulties in designing effective clinical trials to assess progression outcomes. Given that most patients progress slowly, it is generally believed that glaucoma trials require large sample sizes or very long follow-up periods to identify differential rates of disease worsening between treatment groups.⁸

Identifying fast progressors is further complicated by test-retest variability of the commonly available clinical tools, such as standard automated perimetry (SAP) and optical coherence tomography (OCT). Many strategies have been proposed to improve the efficacy of detecting progression in glaucoma. Clustering tests at both ends of the observation period – that is, having multiple measurements at the start and end of the period – has been proposed as a strategy to improve the efficacy of statistical tests to assess progression using trend-based analyses of test measurements over time.^9,10 This approach can more accurately estimate the starting and ending points of the linear regression line, which can improve detection of significant changes over time. Using simulated data, Crabb and colleagues¹⁰ proposed the “wait and see” approach by which clustering tests at the start and end of the observation period would lead to better detection of fast progressors. Other simulation studies have also shown the potential of clustered testing to reduce sample sizes in clinical trials.^11,12 In addition to their potential for leading to more efficient trials, a clustered testing approach may prove beneficial for prompt evaluation of patients whose further progression could place them at high risk of visual disability, severely and irreversibly impacting their quality of life.

While there is a theoretical basis for the efficacy of the clustered testing approach, there is an absence of practical or real-world evidence supporting its benefit in shortening the time to detect glaucoma progression or improving the detection of fast progressors. The United Kingdom Glaucoma Treatment Study (UKGTS) employed clustered testing at baseline, 18 and 24 months in a randomized trial comparing latanoprost versus placebo.¹³ However, although the study identified differences in progression rates between the two groups, it was not designed to assess whether a clustered testing approach would allow for rapid identification of fast progressors. Crucially, the efficiency of the clustering approach could be greatly diminished if significant autocorrelation is present among the tests in a cluster,¹⁴ a hypothesis that could not be tested in prior studies.

In this context, we report the results of the Fast-PACE (Progression Assessment through Clustered Evaluation) study. This investigation was designed to assess the feasibility and efficacy of detecting fast progressors in a short time by intensive clustering of functional and structural testing. Detecting fast progression has major implications, and methods that achieve this could change the paradigm for testing patients clinically and in clinical trials.

METHODS

This was an observational cohort study conducted at the Duke Eye Center, Duke University, Durham, NC. The institutional review board approved the study methodology, which adhered to the tenets of the Declaration of Helsinki and to the Health Insurance Portability and Accountability Act. Informed consent was obtained from all participants.

For inclusion, subjects had to be diagnosed with primary open-angle glaucoma (POAG) at a visit before enrollment. For each participant, both eyes were required to have glaucomatous appearing optic discs on clinical exam and at least one eye had to show evidence of corresponding repeatable (i.e., at least two consecutive) abnormal SAP visual field test results. Abnormal SAP was defined as a pattern standard deviation (PSD) outside of the 95% normal confidence limits, or a Glaucoma Hemifield Test result of “outside normal limits”. Subjects were excluded if they presented with a best-corrected visual acuity less than 20/40, spherical refraction outside ± 5.0 diopters and/or cylinder correction outside 3.0 diopters, or any other ocular or systemic disease that could affect the optic nerve or the visual field. Eyes with a previous history of trabeculectomy or tube shunt surgery were excluded.

Clustered testing

Enrolled POAG subjects underwent 2 clusters of testing, approximately 6 months apart (mean 6.3 ± 1.1 months). Each cluster consisted of 5 visits scheduled at approximately weekly intervals. After the last visit of the baseline cluster patients were scheduled to come back 6 months later for the first visit of the 6-month cluster. At each visit, subjects underwent IOP measurement, 24–2 and 10–2 SAP testing with Swedish Interactive Thresholding Algorithm (SITA) Standard (Humphrey perimeter, Carl-Zeiss Meditec, Inc., Dublin, CA), and spectral-domain OCT testing with the Spectralis SDOCT (Heidelberg Engineering, Dossenheim, Germany) using 3.45-mm peripapillary circle scans. Visual fields with more than 33% false-negative errors or more than 15% false-positive errors were excluded. Visual fields were manually reviewed for artifacts such as lid and rim artifacts, fatigue effects, inappropriate fixation, and evidence that the visual field results were due to a disease other than glaucoma. SD OCT tests were also manually reviewed. Tests with quality score below 15, or with segmentation errors, evidence of decentration or artifacts were also excluded. At each visit, an attempt was made to repeat the test if a poor-quality test was identified.

Participants were treated at the discretion of the attending ophthalmologist. Although changes in treatment were allowed, none occurred during the initial 6-month clustering phase in any subject. After this phase, clinicians saw the study test results and may have adjusted treatments based on these insights, in addition to their routine clinical assessment. This study also includes results from an extended follow-up period post-clustering. Each participant attended a minimum of one visit every six months until the end of follow-up.

Data Analysis

Progression was assessed by trend analyses (rate of change) of SAP 24–2 MD, SAP 10–2 MD and SD OCT global peripapillary RNFL thickness over time, obtained with ordinary least-squares (OLS) linear regression.^9,10 The linear regression slope (rate of change) was calculated based on all 10 tests from the 6-month clustering period. Rates of change were also calculated based on all tests available during the entire follow-up, including cluster and post-cluster tests. For calculations of linear regression slopes, each test result at each visit was considered as an observation, i.e., tests results from each cluster were not averaged together.

For SAP 24–2 MD and SAP 10–2 MD, progression was considered to have occurred if the upper limit of the 95% confidence interval was below 0 (i.e., a statistically significant negative slope). For RNFL thickness, we tested the statistical significance of the RNFL rate of change in relation to the mean estimate of age-related changes (−0.5μm/year), as found in previous studies.¹⁵ Therefore, progression for RNFL was defined to have occurred if the upper limit of the 95% confidence interval was worsening at more than 0.5μm/year. This has been shown to improve specificity of RNFL assessment for detecting progression.¹⁵ Test results from eyes that were determined to have statistically significant change by linear regression were reviewed by two glaucoma specialists, masked to other test results. Confirmed progression was declared if both graders reached a consensus that the changes were likely due to glaucoma deterioration and not the result of artifacts or other disorders.

In our analysis, given that tests within a cluster were obtained within a short timeframe, there is a potential for intra-cluster measurements to be highly correlated. Such correlations might influence the reliability of slopes derived from OLS regression, which assumes uncorrelated residuals. To gauge the extent of this correlation, we computed the intraclass correlation coefficient (ICC) for each eye across the two clusters, for each test. A high ICC would suggest greater consistency within each cluster, raising the concern that even slight variability across the two clusters might lead to exaggerated or spurious slopes in OLS. Conversely, a low ICC, signifying greater variability within clusters than between them, would strengthen the case that a significant slope represents genuine change. This is because despite the high within-cluster variability, a statistically significant trend emerges across clusters, likely indicating true progression rather than fluctuations over the 6-month period. In addition to the ICC analysis, we compared the within-cluster standard deviation between eyes categorized as showing progression and those not showing progression. If the average within-cluster standard deviation is consistent across these groups, it reinforces the argument that significant slopes in the progression group are representative of true disease progression rather than being artifacts.

The required sample size for the study was determined as the number of eyes required to detect a minimum difference of 1 standard deviation (SD) in mean rates of change for SAP 24–2 MD (main outcome) between progressors and non-progressors assuming 15% of eyes would show progression. The calculation also assumed a standard deviation of 0.5 dB/year for the rate of change and a correlation of 0.5 between outcomes of the two eyes of the same subject.¹⁶ The required sample size was established to be 120 eyes for an alpha level of 0.05 and 80% statistical power.

Sensitivity and specificity estimates were calculated assuming progression detected during the overall follow-up as the reference standard. This was done for each test separately and for the combination of tests. We also evaluated the accuracy of the 6-month clustering to identify fast progression during the overall follow-up. For SAP, fast progression was considered as rate of MD change faster than −1dB/year. For global RNFL thickness, fast progression was considered as a rate faster than −2μm/year. We also estimated positive and negative diagnostic likelihood ratios (LRs) as measures of diagnostic accuracy. LRs represent the best way to incorporate diagnostic test results in clinical practice according to the principles of evidence-based medicine.^17,18 The LR for a given test result indicates how much that result will change the probability of disease, going from a pre-test probability (i.e., probability of disease before the test) to a post-test probability. A value of one means that the test provides no additional information, and ratios above or below one increase or decrease the likelihood of disease, respectively. Based on a prior classification definition, LRs greater than 10 or lower than 0.1 would be associated with large effects on post-test probability, LRs from 5 to 10 or from 0.1 to 0.2 would be associated with moderate effects, LRs from 2 to 5 or from 0.2 to 0.5 would be associated with small effects, and LRs closer to one would be insignificant. Confidence intervals for LRs were calculated according to the method proposed by Simel et al.¹⁹

We report median (interquartile range) for rates of change in SAP MD, as the distribution of these rates tend to be highly skewed to the left.^3,5,20,21 For consistency, we also report median rates of RNFL thickness. Mean and standard deviation are reported for other variables with normal distribution. For the main proportions, our estimates are presented along with their 95% confidence intervals. Generalized Estimating Equations (GEE) were employed to adjust for correlations between eyes from the same individual, both in the estimation of these intervals and in hypothesis testing. Statistical analyses were performed using Stata 17 (Statacorp, Inc., College Station, TX). The alpha error was set at 0.05.

RESULTS

6-month Clustering Period

The study included 125 eyes of 65 subjects who completed both the baseline and 6-month clustering. Mean age of subjects at baseline was 68.0 ± 7.5 years; 25% of subjects self-identified as Black, 75% as White, and 58% as female. Median baseline SAP 24–2 MD and SAP 10–2 MD, calculated from the average of the tests in the baseline cluster, were −2.39 (−6.05, −0.45) dB and −1.47 (−4.78, −0.07) dB. Median baseline global RNFL thickness was 73.7 (62.9, 85.9) μm.

The median rate of SAP 24–2 MD change during the 6-month clustering period was −0.33 (−1.79, 0.93) dB/year. Nineteen eyes (15%, CI: 9%−24%) progressed based on SAP 24–2 MD over the 6-month period. Median rates of change for progressors and non-progressors were −2.70 (−3.77, −1.71) dB/year and 0.02 (−1.08, 1.32) dB/year (P<0.001), respectively. Figure 1 shows distributions of rates of 24–2 MD change in progressors versus non-progressors, while Figure 2 shows a scatterplot comparing SAP 24–2 MD values at the baseline and 6-month clusters in the two groups.

Figure 1.

Distributions of rates of change for each one the 3 tests used to assess progression: standard automated perimetry (SAP) mean deviation (MD) 24–2, SAP MD 10–2 and spectral-domain optical coherence tomography global retinal nerve fiber layer (RNFL) thickness.

Figure 2.

Scatterplots illustrating the relationships between measurements from the baseline cluster and the 6-month cluster for each one of the testing modalities, for progressors and non-progressors. Each dot corresponds to a test at equivalent time points in the baseline versus 6-month cluster.

For SAP 10–2 MD, 14 of 125 eyes (11%, CI: 6%−20%) progressed over the 6-month clustering period. Median rates of change for progressors and non-progressors for SAP 10–2 MD were −2.98 (−4.15, −2.30) dB/year and 0.38 (−0.58, 1.19) dB/year (P<0.001). Figure 1 shows the distribution of rates of 10–2 MD in both groups. Most of the eyes progressing by 10–2 (11 out of 14, or 79%) were also identified as progressing by SAP 24–2, while 11 of the 19 eyes (58%) identified as progressing by 24–2 were also identified as progressing by 10–2. There was a strong correlation between rates of change of SAP 24–2 MD and SAP 10–2 MD (r = 0.71; P<0.001; Figure 3). Overall, 22 of the 125 eyes (18%, CI: 10%−27%) had visual field progression over the 6-month clustering period considering 24–2 and 10–2 tests. Table 1 shows demographic and clinic characteristics of eyes that had progressive visual field loss versus those that did not. Mean IOP over the 6-month clustering period was significantly higher in eyes with progressive visual field loss versus those non-progressing (17.6 ± 7.0 mmHg vs. 15.1 ± 3.6 mmHg; P = 0.016). Peak IOP over the same time was also significantly higher in progressing eyes (20.1 ± 8.5 vs. 17.1 ± 4.0 mmHg; P = 0.012), as well as IOP fluctuation (2.2 ± 1.6 vs. 1.6 ± 0.8 mmHg; P = 0.011). Figures 4 and 5 show 24–2 and 10–2 visual fields of example eyes included in the study.

Figure 3.

Scatterplot illustrating the relationship between rates of standard automated perimetry mean deviation (MD) change for 24–2 and 10–2. The zoomed in plot corresponds to the observations from the dashed box on the left.

Table 1.

Demographic and clinical characteristics of eyes that showed visual field progression on either 24-2 or 10-2 standard automated perimetry during the 6-month clustering period.

	Progressor (n = 22)	Non-progressor (n = 103)	P
Age, years	65.3 ± 7.6	68.2 ± 7.4	0.149
Sex, % female	64%	57%	0.702
Race, % Black	32%	23%	0.321
Baseline 24-2 MD, dB  Mean ± SD	−2.75 (−6.56, −0.76) −5.63 ± 7.05	−2.05 (−6.03, −0.16) −4.58 ± 6.92	0.615
Baseline 10-2 MD, dB  Mean ± SD	−1.35 (−4.67, −0.31) −3.82 ± 6.43	−1.56 (−5.73, 0) −4.16 ± 6.58	0.860
Baseline RNFL thickness, μm  Mean ± SD	67.6 (58.8, 74.4) 65.9 ± 13.3	75.3 (63.6, 87.7) 76.6 ± 18.9	0.011
Rate SAP 24-2 MD (dB/year)  Mean ± SD	−2.42 (−3.51, −1.53) −3.18 ± 3.68	−0.004 (−1.08, 1.32) −0.05 ± 1.85	<0.001
Rate SAP 10-2 MD (dB/year)  Mean ± SD	−2.38 (−3.95, −1.04) −3.05 ± 3.35	0.42 (−0.46, 1.20) 0.40 ± 1.73	<0.001
Rate RNFL Thickness, μm/year  Mean ± SD	−0.40 (−0.96, 1.10) −0.17 ± 1.44	−0.33 (−2.40, 0.90) −0.58 ± 3.49	0.604
Mean IOP, mmHg	17.6 ± 7.0	15.1 ± 3.6	0.016
Peak IOP, mmHg	20.1 ± 8.5	17.1 ± 4.0	0.012
IOP fluctuation, mmHg	2.21 ± 1.63	1.61 ± 0.77	0.011
CCT (μm)	541 ± 49	536 ± 42	0.656

^*MD: mean deviation, IOP: intraocular pressure, CCT: central corneal thickness, RNFL: retinal nerve fiber layer. Boldface indicates statistical significance (P<0.05).

Figure 4.

Greyscale and pattern deviation plots from standard automated perimetry (SAP) 24–2 (top rows) and 10–2 (bottom rows) tests for an eye with early glaucoma at baseline. The 5 tests for each cluster are shown (baseline and 6-month), along with the last test available during follow-up. The scatterplots below show the rates of change calculated from just the observations from the two clusters (6-month clustering, black line) as well as the rate of change calculated from all observations available during follow-up (overall follow-up, grey line).

Figure 5.

Greyscale and pattern deviation plots from standard automated perimetry (SAP) 24–2 (top rows) and 10–2 (bottom rows) tests for an eye with severe glaucoma at baseline. The 5 tests for each cluster are shown (baseline and 6-month), along with the last test available during follow-up. The scatterplots below show the rates of change calculated from just the observations from the two clusters (6-month clustering, black line) as well as the rate of change calculated from all observations available during follow-up (overall follow-up, grey line). The eye had a trabeculectomy at 6 months of follow-up. It is possible to see that the measurements stabilize after that. The rate during the initial 6-month clustering period is much more negative than the overall rate.

Sixteen eyes (13%, CI: 8%−21%) progressed by global RNFL thickness. Median rates of change for progressors and non-progressors were −2.43 (−4.02, −0.98) μm/year and −0.18 (−1.75, 1.19) μm/year (P<0.001). Figure 1 also shows distributions of rates of global RNFL change in both groups, while Figure 2 shows a scatterplot comparing measurements at the baseline and 6-month clusters. Table 2 shows demographic and clinic characteristics of eyes that had progressive RNFL loss versus those that did not. Mean IOP over the 6-month clustering period was significantly higher in progressing versus non-progressing eyes (18.8 ± 6.2 mmHg vs. 15.1 ± 4.0 mmHg; P=0.001). Peak IOP was also significantly higher in progressing eyes (21.8 ± 7.8 vs. 17.0 ± 4.4 mmHg; P<0.001), as well as IOP fluctuation (2.6 ± 1.7 vs. 1.6 ± 0.8 mmHg; P<0.001). Figure 6 shows an example of a representative eye that exhibited RNFL progression during the study.

Table 2.

Demographic and clinical characteristics of eyes that showed progression on spectral-domain optical coherence tomography (SD OCT) global retinal nerve fiber layer thickness during the 6-month clustering period.

	Progressor (n = 16)	Non-progressor (n = 109)	P
Age, years	66.4 ± 6.7	68.0 ± 7.6	0.496
Sex, % female	61%	56%	0.773
Race, % Black	13%	27%	0.392
Baseline 24-2 MD, dB  Mean ± SD	−2.09 (−5.79, −0.83) −4.17 ± 5.37	−2.39 (−6.10, −0.17) −4.85 ± 7.14	0.689
Baseline 10-2 MD, dB  Mean ± SD	−1.20 (−4.21, −0.51) −1.92 ± 2.02	−1.56 (−6.13, 0.06) −4.42 ± 6.89	0.007
Baseline RNFL thickness, μm  Mean ± SD	71.7 (59.7, 83.4) 71.9 ± 16.2	74.1 (63.2, 86.3) 75.2 ± 18.8	0.536
Rate SAP 24-2 MD (dB/year) Mean ± SD	−0.92 (−2.67, 0.16) −2.36 ± 4.61	−0.16 (−1.72, 1.16) −0.35 ± 2.01	0.003
Rate SAP 10-2 MD (dB/year) Mean ± SD	−0.53 (−2.41, 1.07) −1.75 ± 3.87	0.27 (−0.95, 1.02) 0.02 ± 2.13	0.007
Rate RNFL Thickness, μm/year Mean ± SD	−2.43 (−4.02, −0.98) −2.96 ± 2.42	−0.18 (−1.75, 1.19) −0.15 ± 3.19	<0.001
Mean IOP, mmHg	18.8 ± 6.2	15.1 ± 4.0	0.001
Peak IOP, mmHg	21.8 ± 7.8	17.0 ± 4.4	<0.001
IOP fluctuation, mmHg	2.60 ± 1.74	1.59 ± 0.77	0.001
CCT (μm)	542 ± 45	536 ± 43	0.611

MD: mean deviation, IOP: intraocular pressure, CCT: central corneal thickness, RNFL: retinal nerve fiber layer. Boldface indicates statistical significance (P<0.05).

Figure 6.

Spectral-domain optical coherence tomography scans obtained during the baseline and 6-month cluster for one of the eyes in the study. The last follow-up scan is shown in the bottom. It is possible to see thinning of the retinal nerve fiber layer (RNFL) thickness inferiorly. It is possible to see that the rate during the 6-month clustering period (black line) is approximately the same as the overall follow-up rate (grey line) calculated from all observations during follow-up.

Of the 16 eyes progressing by global RNFL rates, 8 (50%) were also detected as progressing by visual fields. Figure S7 (available at https://www.aaojournal.org) shows a Venn diagram illustrating progression as detected by each test during the 6-month clustering period. A total of 30 of 125 eyes (24%, CI: 16%−34%) progressed by either SAP (24–2 or 10–2) or SD OCT rates over the 6-month clustering period, whereas 95 (76%) eyes showed no progression on these tests. A total of 22 out of the 65 POAG subjects (34%) had at least 1 eye with progression by either structure or function. 8 out of 65 POAG subjects had bilateral progression (12%).

Accuracy of 6-month Clustering in Predicting Progression During Extended Follow-up

The 125 eyes included in the study had an overall mean follow-up time of 25 ± 16 months with a mean of 17 ± 6 visits. The median rate of SAP 24–2 MD change was −0.04 (−0.53, 0.37) dB/year for all 125 eyes. A total of 24 out of 125 eyes (19%) showed progression on SAP 24–2 MD during follow-up with median rates of change for progressors and non-progressors of −0.64 (−1.37, −0.41) dB/year versus 0.11 (−0.29, 0.61) dB/year (P<0.001). For SAP 10–2, 19 eyes (15%) progressed, with median rates of change for progressors and non-progressors of −0.82 (−2.41, −0.42) and 0.16 (−0.12, 0.53) dB/year (P<0.001). For SD OCT, 24 out of 125 eyes (19%) progressed, with median rates of change for progressors and non-progressors of −1.61 (−2.19, −1.28) μm/year and −0.18 (−0.86, 0.48) (P<0.001).

Table 3 summarizes the accuracy of the 6-month clustering to identify progression, when the overall follow-up for each corresponding test is assumed as the reference standard. It also shows the accuracy for detecting progression based on the overall combination of tests. 35 eyes (28%) progressed during the overall follow-up period on SAP 24–2, 10–2 or SD OCT. The 6-month clustering detected 25 eyes as progressing, for a sensitivity of 71%, (CI: 53%−85%). Of the 90 eyes that did not progress during the overall follow-up, 85 were also declared as non-progressing during the 6-month period, for a specificity of 94% (CI: 88% – 98%). The LR+ was 12.9 (CI: 7.0 – 23.6) and LR− was 0.30 (CI: 0.21 – 0.43).

Table 3.

Diagnostic accuracy for the 6-month clustering approach when the overall follow-up for each corresponding test (or combination of tests) is assumed as the reference standard.

Test	Sensitivity (95% CI)	Specificity (95% CI)	LR+ (95% CI)	LR− (95% CI)
SAP 24-2 MD	58% (37% – 78%)	95% (89% – 98%)	11.8 (8.4 – 16.6)	0.44 (0.31 – 0.62)
SAP 10-2 MD	37% (16% – 62%)	93% (87% – 97%)	5.3 (2.31 – 13.1)	0.68 (0.52 – 0.88)
SDOCT global RNFL thickness	58% (37% – 78%)	98% (93% – 100%)	29.8 (10.3 – 86.4)	0.43 (0.31 – 0.59)
All tests	71%, (53%–85%)	94% (88% – 98%)	12.9 (7.0 – 23.6)	0.30 (0.21 – 0.43)

^*MD: mean deviation, RNFL: retinal nerve fiber layer, LR: likelihood ratio, CI: confidence interval

Accuracy of 6-month Clustering in Predicting Fast Progressors

Table 4 summarizes the accuracy of the 6-month clustering to identify fast progression during the overall follow-up. From the 125 eyes, 14 (11%) were considered fast progressors by SAP 24–2, 10–2 or SD OCT over the whole duration of follow-up for each eye. The 6-month clustering was able to identify 13 of these eyes as progressing, for a sensitivity of 93% (CI: 66% – 100%). Only 1 eye with fast progression during the overall follow-up was not detected as a fast progressor during the initial 6-month clustering period. This eye had fast RNFL progression on SD OCT which was not detected during the 6-month clustering period, but visual fields remained unchanged during the overall follow-up. Of the 111 eyes that were not fast progressors during overall follow-up, 94 were not progressing by the clustering method, resulting in a specificity of 85% (CI: 77% – 90%). Using fast progression as reference standard, LR+ was 6.0 (CI: 4.2 – 8.7) and LR− was 0.08 (CI: 0.01 – 0.59).

Table 4.

Diagnostic accuracy for the 6-month clustering approach when fast progression* in the overall follow-up for each corresponding test (or combination of tests) is assumed as the reference standard.

Test	Sensitivity (95% CI)	Specificity (95% CI)	LR+ (95% CI)	LR− (95% CI)
SAP 24-2 MD	88% (47% – 100%)	90% (83% – 95%)	8.5 (2.9 – 24.7)	0.14 (0.04 – 0.54)
SAP 10-2 MD	75% (35% – 97%)	93% (88% – 97%)	11.0 (4.6 – 24.5)	0.27 (0.11 – 0.66)
SDOCT global RNFL thickness	83% (33% – 99%)	91% (85% – 95%)	9.0 (4.4 – 18.5)	0.18 (0.3 – 1.34)
All tests	93%, (66% – 100%)	85% (77% – 90%)	6.0 (4.2 – 8.7)	0.08 (0.01 – 0.59)

^*Fast progression defined as slope of SAP 24-2 MD or 10-2 MD faster than −1dB/year, or slope of SDOCT global RNFL thickness faster than −2μm/year

MD: mean deviation, RNFL: retinal nerve fiber layer, LR: likelihood ratio, CI: confidence interval

Analysis of within-cluster test correlations

Mean ICC values across the two clusters were 0.15 ± 0.21, 0.13 ± 0.22 and 0.18 ± 0.27 for SAP 24–2 MD, SAP 10–2 MD and SD OCT RNFL thickness, respectively. We also compared the average within-cluster standard deviation for eyes declared as progressing versus non-progressing for each test. For SAP 24–2 MD, average within-cluster standard deviation was 1.29 ± 0.96 dB for progressors and 1.06 ± 0.56 dB for non-progressors (P = 0.426). For SAP 10–2 MD, corresponding values were 1.30 ± 0.81 dB and 0.90 ± 0.68 dB, respectively (P = 0.064). For SD OCT RNFL, corresponding values were 1.32 ± 0.88 μm and 1.16 ± 0.78 μm, respectively (P = 0.513).

Effect of number of tests in a cluster

We also explored the effectiveness of conducting fewer tests within a cluster to detect progression within 6 months. When reduced to 4, 3, or 2 tests per cluster, we found that 74%, 53%, and 21%, respectively, of eyes flagged by the 5-test clusters as progressing were similarly identified.

DISCUSSION

An intensive, clustered testing approach detected a substantial proportion of eyes experiencing rapid glaucoma progression within a short time interval. By administering a five-test cluster at baseline and another set six months later, we identified approximately a quarter of the tested eyes as progressing. This was determined through either functional or structural assessment, employing conventional conservative criteria based on the rates of change in global summary metrics. Most of these eyes were confirmed as progressing based on the extended follow-up. The implications could be significant, especially for how best to monitor high-risk eyes prone to vision loss and how to design clinical trials aimed at investigating therapies to slow down disease deterioration.

In this study involving 125 eyes with POAG, nearly one in five eyes progressed based on 24–2 or 10–2 testing over only six months of observation. A high correlation was observed between the rates of change on both tests, reinforcing the validity of the findings. It is improbable that such changes in MD would have resulted from media opacities or cataracts over only six months. Visual fields were also reviewed to rule out significant artifacts or non-glaucomatous defects as the explanation for the findings. Three eyes that were initially flagged with statistically significant slopes on SAP 24–2 were deemed as non-progressing by the masked graders. A significant proportion of eyes also progressed by RNFL measured using SD OCT. Structural and functional progression often are not simultaneously detected due to factors such as test characteristics, variability, and the inherent structure-function relationship in glaucoma.^22–24 Indeed, only half of eyes progressing by RNFL also progressed by visual field concurrently, a finding consistent with prior studies.^5,22,24 Overall, nearly a quarter of the study population progressed over six months by either visual fields or RNFL.

A key step in validating the 6-month clustering results was to assess its accuracy in relation to the reference standard of progression based on the extended follow-up. Across all tests, we observed a high specificity, close to 95%, indicating a small number of false-positives for the 6-month clustering approach. Sensitivities for identifying progressors varied, with a range from 37% for the 10–2 test to 58% for both the 24–2 and OCT tests (Table 3). Such moderate sensitivity is anticipated because a 6-month window may not capture all slow-progressing eyes. However, a combination of the three tests had a sensitivity to 71% for detecting all cases of progression by structure and/or function during the overall follow-up, while still maintaining a high specificity of 94%. Although some slow progressors are missed in this short timeframe, the impact is likely minimal as these cases would likely be flagged with extended follow-up before major loss of vision occurs. Notably, the 6-month clustering method detected fast progressors with a high sensitivity (93%). This is a key result as identifying fast-progressing eyes at the outset of care very well could lead to more rapid escalation of treatment which should reduce glaucoma-related disability and blindness. Unfortunately, blindness from glaucoma is still relatively common, even in treated patients. In a report by Peters and colleagues,²⁵ 42% of open-angle glaucoma patients had at least 1 eye blind from glaucoma and 16% developed bilateral blindness over a median disease duration of just 12 years.

The calculated diagnostic LRs offer additional insights into the interpretation of the results. The LR+ represents the odds of a positive test result at six months in a progressing eye (as determined at final follow-up) compared to a non-progressing one. An LR+ of 12.9, as observed in our study, implies that an eye with a significant finding on our 6-month clustering method is almost 13 times more likely to be truly progressing during overall follow-up than one without. On the other hand, the LR− quantifies the odds of a negative test result in progressing versus non-progressing eyes. The observed LR− of 0.30 suggests that while a non-significant result in our 6-month assessment considerably reduces the probability of progression, it does not entirely rule it out. When applied to fast progressors, the remarkably low LR− of 0.08 accentuates the test’s strength. It signifies that the absence of progression during the cluster period strongly indicates a low probability of an eye being a fast progressor during the extended follow-up. This lends substantial confidence in ruling out fast progression when the 6-month clustered tests are non-significant.

The significant and clinically meaningful changes we observed over just six months in our study challenge the traditional notion that long periods of observation are necessary to detect glaucoma progression. Visual field test-retest variability often masks genuine worsening of glaucoma. Prior research has indicated that it typically takes around 4 years of annual visual field testing, or 3 years of semi-annual testing, to detect eyes progressing at rates of MD change of −1dB/year.^26,27 In contrast, our intensive, clustered testing approach identified fast progressors in just 6 months. In fact, the short time between clusters reduces the likelihood that this method will successfully identify slow-progressing cases. In OLS linear regression, the standard error (SE) of the estimated rate is tied to the variance of the independent variable, which is time in our case. A larger spread in time (i.e., longer follow-up) results in a smaller SE, making even modest rates potentially statistically significant. Conversely, a small spread in time requires a large rate for statistical significance. This implies that short, 6-month follow-ups are more likely to miss slowly progressing eyes but are effective at identifying rapid progressors. Our results confirm that only one eye with fast progression would have been missed by the 6-month clustered testing, highlighting its utility in catching those at high risk of irreversible vision loss.

The fact that the tests in a cluster were performed at relatively short intervals between them may bring concerns related to excessive within-cluster test correlations. This could potentially lead to problems in estimating regression slopes and their statistical significance. In addition, if the tests in a cluster are too correlated, even small fluctuations in performance between the two clusters could generate spurious slopes deemed statistically significant. However, the small ICCs obtained in our study show that the within-cluster variability generally exceeded the between-cluster variability, suggesting that a statistically significant slope in the 6-month period was likely not triggered by some small fluctuation in performance. We also compared the within-cluster standard deviations and found no difference between eyes declared and progressing versus not. Of course, the main indicator of reliability of the 6-month findings should still be its accuracy in detecting changes as seen in the overall follow-up, as indicated above.

Prior studies generally report slow average rates of glaucoma progression when monitoring patients over extended time periods. Chauhan and colleagues³ reported a median rate of change of −0.05 (0.13, −0.30) dB/year in a large patient cohort under routine care. In a large cohort study with 6138 eyes of 3669 patients, mean rate of MD change was −0.20 dB/year for POAG subjects.⁵ Both studies identified fewer than 5% of eyes as fast progressors. These results seem at odds with the current findings, and with previously reported rates of blindness from glaucoma. These rates are based on long-term linear trends that ignore changes in treatment. Clinicians often adjust treatment based on test results, making those long-term linear trend estimates less accurate for capturing rapid initial changes that triggered intervention. For instance, Figure 5 shows an example of an eye in our study that experienced significant deterioration over 6 months, prompting a trabeculectomy that stabilized the visual fields. If assessed over the entire follow-up, the rate of change would seem much slower. That is why the extended follow-up data of our sample are consistent with earlier studies, with a median rate of change for SAP 24 MD of −0.04 dB/year and a mean rate of −0.15 dB/year for the 125 eyes monitored. Importantly, given the effects of treatment after the cluster period, the design of our study does not allow for a direct comparison of the magnitude of the 6-month slopes to those obtained in the overall follow-up. It is possible that some 6-month slopes may be over- or underestimated due to cluster effects. However, we do not believe that this has significantly impacted estimates of sensitivity and specificity. Although treatment may have been intensified in those eyes detected as progressing in 6 months, it is still expected that an evaluation of all tests over the whole follow-up would still show progression if the 6-month progression was true. Even if one were to argue that intensification of treatment led to slowing of slopes and lack of statistical significance when measured over the long-term, this would have resulted in a bias toward worse performance of the 6-month clustering (i.e., increase in false-positives), rather than spuriously increasing performance.

The impact of treatment can also distort risk factor analysis in observational studies. For example, many studies did not find a link between mean IOP and progression, likely due to treatment interventions lowering IOP in progressing eyes. Over time, this can mask or even reverse the IOP-progression relationship.²⁸ Such treatment effects also compromise the reliability of predictive models developed from such observational data. In contrast, our study’s design enabled the clear identification of higher mean IOP, peak IOP and IOP fluctuation in eyes progressing over the 6-month period, as these eyes did not undergo additional interventions during this period (Tables 1 and 2). Of note, the large overlap in the distributions of IOPs in the progressors and non-progressors shows that although the means of each group were significantly different, it would not be possible to clearly separate them solely based on IOP. Using our clustering framework, future studies could develop more clinically relevant models for predicting risk of glaucoma progression.

The findings of our work are of special relevance for the design of clinical trials investigating new treatments aimed at slowing glaucoma progression, such as neuroprotective therapies. Recent simulations suggest that using rates of MD change as endpoints and employing clustered testing schemes could substantially reduce necessary sample sizes and predict conventional endpoints.^11,29 Our findings empirically confirm the feasibility of identifying progressors within a short time span. However, it is important to recognize that while our 6-month observational period was effective for identifying progressors, Phase III clinical trials may necessitate longer observation periods to comprehensively assess the safety profile of drugs. A short 6-month period could also be useful in early-phase trials to gather preliminary efficacy data, aiding in the planning of Phase III trials.

While the intensive 5-test cluster approach used in our study may be suitable for clinical trials, it may be burdensome for clinical practice. Indeed, most patients may not require this rigorous testing regimen. However, this focused approach could be advantageous for those at a high risk of progression that could lead to significant, irreversible impact on their quality of life. This includes patients with severe visual field loss, those with moderate damage in the better-seeing eye, or those with visual field defects threatening central vision, for example. Alternatively, newly diagnosed patients on stable treatment could undergo this intensive testing to help classify them at the start of care.

Our study has some limitations. Due to the large scale of testing and constrained resources, the sample size was kept relatively small for this initial investigation. However, the sample size allowed for enough power to detect significant differences, as noted. Future studies should be planned to provide more precise confidence intervals for measures of accuracy, though. As another limitation, only summary global metrics were used to assess progression. Although this may have resulted in missing some eyes with localized progression, there is no agreed consensus on how to integrate all information from the many visual field and OCT parameters. Global metrics provide a simple measure of rate of change and have been widely demonstrated to predict disability in glaucoma.^30–33 The lack of a perfect reference standard for glaucoma progression also limits the evaluation of accuracy of the 6-month clustering approach. We used the slopes over the extended follow-up as the reference standard, but other approaches could be considered. Some authors have suggested that “significant improvement” could be used as a proxy for specificity.³⁴ The proportions of statistically significant positive slopes were 7%, 6%, and 7% for SAP 24–2, SAP 10–2, and RNFL, respectively, resulting in estimated specificities of 93%, 94%, and 93%. These numbers are similar to those obtained using the extended follow-up as the reference standard. Of note, this approach does not consider potential learning effects and may underestimate specificity. Finally, it is possible that subjects may have been more inclined to participate in this intensive testing study if they felt that their disease could be getting worse. This selection bias may have resulted in a higher proportion of progressors compared to a typical clinical population. Nonetheless, this would not diminish the clinical relevance of our findings, as our study would still show the potential of the method to identify true progression in suspected cases.

In summary, our intensive clustered testing methodology offers promising implications for clinical trials investigating interventions to slow down glaucoma progression. It may also be of value for short term assessment of high-risk subjects. It is generally believed that protracted observational periods are needed to detect glaucoma progression. However, our findings challenge this, demonstrating the potential of our approach to identify rapid progressors within just six months.

Abbreviations:

IOP

intraocular pressure

IQR

interquartile range

MD

mean deviation

OCT

optical coherence tomography

OLS

ordinary least square

PACE

Progression Assessment through Clustered Evaluation

POAG

primary open angle glaucoma

PSD

pattern standard deviation

RNFL

retinal nerve fiber layer

SAP

standard automated perimetry

SD

spectral domain

SE

standard error

SITA

Swedish Interactive Thresholding Algorithm