SCORE2 Report 24: Nonlinear Relationship of Retinal Thickness and Visual Acuity in Central Retinal and Hemiretinal Vein Occlusion

Abstract

Purpose:

To investigate whether a nonlinear association between central subfield thickness (CST) on spectral domain optical coherence tomography (OCT) and concurrent visual acuity letter score (VALS) exists in eyes treated initially with aflibercept or bevacizumab for macular edema associated with central retinal (CRVO) or hemiretinal vein occlusion (HRVO) in the Study of COmparative Treatments for REtinal Vein Occlusion 2 (SCORE2).

Design:

Long-term follow-up after a randomized clinical trial from sixty-four centers in the United States.

Participants:

Participants were followed up to 60 Months and treated at investigator discretion after completing the 12-month treatment protocol.

Methods:

Two-segment linear regression models were compared to simple linear regression models of VALS on CST. Pearson correlation coefficients were calculated to assess strength of CST-VALS associations.

Main Outcome Measures:

CST was measured by OCT and VALS by the electronic Early Treatment Diabetic Retinopathy Study (e-ETDRS) methodology.

Results:

Estimated inflection points, reflecting a turning point at which the CST-VALS association changes from positive to negative, calculated at 7 post-baseline visits, range from 217-256 microns. There is a strongly positive correlation to the left of each estimated inflection point, ranging from 0.29 (P<0.01 at Month 60) to 0.50 (P<0.01 at Month 12), and a strongly negative correlation to the right of each estimated inflection point, ranging from −0.43 (P<0.01 at Month 1) to −0.74 (P<0.01 at Month 24). Randomization statistical tests show that 2-segment models are favored over 1-segment models for all post-baseline months (P<0.001 for all tests performed).

Conclusions:

The relationship between CST and VALS in eyes with CRVO or HRVO after treatment with anti-vascular endothelial growth factor (VEGF) therapy is not simply linear. The usually modest correlations between OCT-measured central retinal thickness and visual acuity belie strong left and right correlations present in 2-segment models. The SCORE2 participants with post-treatment CST close to the estimated inflection points of 217 to 256 μm showed the best VALS. In patients treated with anti-VEGF for macular edema associated with CRVO or HRVO, a thinner retina is not always associated with better VALS.

Trial Registration:

Clinical trial identifier at clinicaltrials.gov: NCT01969708.

Précis

We found a nonlinear relationship between central subfield thickness and visual acuity after anti-VEGF treatment in central retinal or hemiretinal vein occlusion eyes.

Introduction

Since it became commercially available in 1995, optical coherence tomography (OCT) has provided useful information on vitreoretinal morphological changes associated with a variety of posterior segment diseases.¹ OCT has been used to assess the outcomes of treatment for macular edema associated with myriad posterior segment diseases, such as diabetic macular edema, branch retinal vein occlusion (BRVO), and central retinal vein occlusion (CRVO).^2-27

Treatments for macular edema, such as photocoagulation, intravitreal steroids, and intravitreal anti-vascular endothelial growth factor therapy, are targeted to reduce retinal thickness, with the assumption that a reduction in retinal thickness will be associated with an improvement in visual acuity. Published studies have reported an inverse relationship between OCT-measured central retinal thickness and visual acuity, with correlation coefficients, either reported or calculated from the literature reports, ranging from a magnitude of 0.16 to 0.64 among patients with diabetic macular edema.^21-27 In the Standard Care versus COrticosteroid for REtinal Vein Occlusion (SCORE) Study, the correlation coefficient for the association between baseline OCT-measured center point thickness and best-corrected electronic Early Treatment Diabetic Retinopathy Study (e-ETDRS) visual acuity letter score (VALS) was −0.27 (95% confidence limits: −0.38, −0.16) for participants in the CRVO trial and −0.28 (95% confidence limits: −0.37, −0.19) in the BRVO trial.²⁸ Ciulla et al.²⁹, in an analysis of 6 prospective clinical trials (4 of which were randomized and all of which used standardized assessment for both OCT acquisition and VALS measurement) also reported poor to modest associations between baseline OCT-measured central subfield thickness (CST) and best-corrected ETDRS VALS (−0.56 in patients with macular edema from retinal vein occlusion; −0.50 in patients with diabetic macular edema, and −0.38 in patients with noninfectious uveitis).

A linear relationship, however, cannot fully describe the association between retinal thickness and visual acuity, since a retina that is too thin likely cannot support good vision. The purpose of the current study is, among participants treated initially with aflibercept (EYLEA^®) or bevacizumab (AVASTIN^®) for macular edema associated with CRVO or hemiretinal vein occlusion (HRVO) in the Study of COmparative Treatments for REtinal Vein Occlusion 2 (SCORE2), to more fully investigate the relationship between OCT-measured central retinal thickness and visual acuity. Specifically, we explore whether a nonlinear relationship between post-baseline CST and concurrent VALS after treatment in SCORE2 describes this association better than the poor to modest linear relationship that has been demonstrated in many prior studies.

Methods

The SCORE2 Study design and methods, described in detail elsewhere,³⁰ are summarized here. The study adhered to the tenets of the Declaration of Helsinki³¹ and is registered on http://www.clinicaltrials.gov (identifier: NCT01969708). Institutional review board (IRB) approval of the protocol was obtained, and written informed consent was obtained from all participants.

Between September 17, 2014 and November 18, 2015, a total of 362 patients (305 with CRVO and 57 with HRVO) were randomly assigned to receive intravitreal injection of bevacizumab (1.25 mg) or aflibercept (2.0 mg) at randomization and every 4 weeks through Month 5. The primary outcome was change from baseline in best-corrected e-ETDRS VALS at Month 6, with a non-inferiority margin of 5.^18,29 Following assessment of the primary outcome at Month 6, participants originally assigned to aflibercept who met the protocol-defined criteria for a good response were re-randomized to either continuing aflibercept every 4 weeks (n=79) versus changing to a treat and extend (TAE) regimen (n=80); participants originally assigned to aflibercept who met the protocol-defined criteria for a poor or marginal response at 6 months (n=15) were to receive a dexamethasone implant. Participants originally assigned to bevacizumab who met the protocol-defined criteria for a good response were re-randomized to either continuing bevacizumab every 4 weeks (n=67) versus changing to a TAE regimen (n=67); participants originally assigned to bevacizumab who met the protocol-defined criteria for a poor or marginal response at 6 months (n=39) were to receive aflibercept. SCORE2 participants’ last visit as part of the SCORE2 protocol-defined treatment schedule was at Month 12.

After Month 12, there was no protocol-defined treatment schedule. Rather, physicians could treat as they deemed necessary, using any commercially available drug (including non-study drug or no drug) based on their typical practice and on any schedule, and patients were followed at visits through Month 60 as part of the SCORE2 Long-term Follow-up (SCORE2 LTF).^19,20 Study data included all interventions administered to the study eye for treatment of macular edema secondary to CRVO or HRVO (including injections given at non-study offices, provided they were documented in the medical record). At Months 0, 6, 12, 24, 36, 48, and 60, data were collected on best-corrected e-ETDRS VALS, central subfield thickness (CST) assessed by spectral domain OCT (SD-OCT), and eye examinations. SD-OCT images were sent to the Reading Center at the University of Wisconsin-Madison for grading. To standardize CST measurements across OCT manufacturers, the Reading Center dicomized and segmented all OCT scans. CST thickness was measured from the top of the inner limiting membrane to the top of the retinal pigment epithelium. At each annual in-person visit with the participant, and via telephone call with the site clinical coordinator at Months 18, 30, 42, and 54, there was a medical record review of new ocular conditions, procedures, and other new health conditions occurring since the preceding annual visit. Data for these analyses were frozen on May 6, 2021.

Statistical Analyses

Graphical analyses included a scatterplot of VALS on the vertical axis versus CST on the horizontal axis for a set of study eyes from SCORE2 participants at a specific visit. We regressed VALS on CST, and further investigated whether a two-segment linear model significantly improved the fit over a simple linear regression line. The two-segment model has an inflection point at some CST value, say $CST=K$, and two straight lines, one to the left of $K(CST\le K)$, and the other to the right ($CST\ge K$). The two lines have different slopes and meet (that is, have the same VALS value) when $CST=K$. The two-segment model, which, when $K$ is specified, can be fit via ordinary least squares, is:

$$VAL{S}_i=\alpha +\beta CS{T}_i+\gamma (CS{T}_i-K{)}_{+}+{ϵ}_i$$

where $(u{)}_{+}=max(0,u)$.

It turns out that $\beta$ and ($\beta +\gamma$) are the values of the left and right slopes, respectively. Because $\gamma$ is the difference between slopes, testing whether $\gamma$ differs significantly from 0 reveals whether the two-segment model is significantly better than the simple one-segment regression of VALS on CST, but the test must be adjusted because $K$ must be estimated. We estimate $K$ to be the observed CST value that minimizes the sum of squared errors (SSE). An appendix shows the Randomization test we used to adjust the $\gamma$ test. We also analyzed the data via the R Segmented package,³² which provides an adjusted test of $\gamma$ via a pseudo Score test.³³ Our Randomization approach agreed well with the results of the Segmented package. We also used the Segmented package to test whether more than 2 segments were necessary. Our graph compares the 2-segment model to locally estimated scatterplot smoothing (LOESS).³⁴ LOESS is a popular smoothing method with no preconceptions about global linearity or the number of segments. We used the SAS PROC LOESS defaults throughout. All analyses were carried out in SAS 9.4 and R version 3.6.2, Segmented package version 1.6.0. We calculate the CST-VALS correlation for the two-segment regression model as the Pearson correlation between observed VALS and the corresponding VALS value predicted by the two-segment regression.

Results

Scatterplots of VALS on concurrent SD-OCT CST for all available study eyes from SCORE2 participants at baseline (before treatment in SCORE2) and at Months 1, 6, 12, 24, 36, 48, and 60 are presented within the panels of Figure 1. Estimated inflection points, reflecting a turning point at which the CST-VALS association changes from positive to negative, calculated for each of the 7 post-baseline visits, are marked by vertical reference lines, and range from 217 microns to 256 microns. The association of VALS and CST can be visually assessed on each panel of Figure 1 through both the LOESS fit (solid line segments) and the fit from the two-segment linear model (dashed line segments). There is striking agreement between the 2-segment regressions and the LOESS lines in every panel of the graph.

Figure 1.

Scatterplots of VALS against concurrent CST for all available SCORE2 participants at baseline (before treatment) and Months 1, 6, 12, 24, 36, 48, and 60. Vertical reference lines mark estimated inflection points. Each panel also shows a LOESS fit (solid line segments) and the fit from the two-segment linear model (dashed line segments).

Columns 1-4 of Table 1 display Pearson correlations for (1) the data as a whole, (2) the left side ($CST\le K$), (3) the right side ($CST\ge K$), (4) the two-segment model. The numerical value of the estimated inflection point $K$ is shown in column 5. Inflection points were also calculated for Months 2-5 and 7-11, but lie within the range of post-baseline inflection points reported in the text and shown in Table 1, and are omitted for the sake of concision. Considering post-baseline visits, there is a strongly positive correlation to the left of the estimated inflection point (column 2), ranging from 0.29 (P<0.01 at Month 60) to 0.50 (P<0.01 at Month 12), and a strongly negative correlation to the right of the estimated inflection point (column 3), ranging from −0.43 (P<0.01 at Month 1) to −0.74 (P<0.01 at Month 24). The negative post-baseline overall (one-segment) correlations are weaker than the strong negative right and, except for Months 24 and 48, strong positive left correlations. Note that the correlation p-values are not adjusted to account for estimation of the inflection point. The 2-segment correlation (column 4) exceeds the 1-segment correlation (column 1) in magnitude by 0.15 to 0.38 correlation units across the post-baseline visits. At baseline, the 2-segment approach seems unnecessary, perhaps because most CST values exceed the inflection point values shown at later visits. The patterns and corresponding correlation coefficients examined separately for those initially randomized to aflibercept are similar to those initially randomized to bevacizumab, so we present the treatment groups combined in Figure 1.

Table 1:

Visual Acuity and Concurrent Central Subfield Thickness at Selected Visits

Visit	Correlation coefficient (P-value)*				Inflection Point (microns) with 95% Confidence Intervals	Randomization test p-values	Pseudo score test P-values
	Linear relationship	2- segment model – left side	2- segment model – right side	2- segment model – overall			2nd segment	3rd segment
Column	1	2	3	4	5	6	7	8
Baseline	−0.47 (P<0.01)	−0.09 (P=0.32)	−0.45 (P<0.01)	0.48 (P<0.01)	559 (318, 800)	0.56	0.22	0.84
Month 1	−0.20 (P<0.01)	0.43 (P<0.01)	−0.43 (P<0.01)	0.43 (P<0.01)	256 (239, 272)	<0.001	0.002	0.01
Month 6	−0.14 (P=0.01)	0.49 (P<0.01)	−0.52 (P<0.01)	0.52 (P<0.01)	227 (217, 238)	<0.001	<.0001	0.22
Month 12	−0.16 (P<0.01)	0.50 (P<0.01)	−0.45 (P<0.01)	0.49 (P<0.01)	224 (212, 236)	<0.001	<.0001	0.25
Month 24	−0.51 (P<0.01)	0.45 (P<0.01)	−0.74 (P<0.01)	0.66 (P<0.01)	239 (222, 255)	<0.001	.001	0.17
Month 36	−0.33 (P<0.01)	0.49 (P<0.01)	−0.54 (P<0.01)	0.52 (P<0.01)	222 (203, 241)	<0.001	0.02	0.10
Month 48	−0.38 (P<0.01)	0.37 (P<0.01)	−0.53 (P<0.01)	0.49 (P<0.01)	217 (192, 242)	<0.001	0.52	0.06
Month 60	−0.29 (P<0.01)	0.29 (P=0.01)	−0.55 (P<0.01)	0.44 (P<0.01)	235 (205, 265)	<0.001	0.06	0.55

Columns 1-4 show correlations for (1) the simple linear model (2) the left side (3) the right side (4) the 2-segment model. No p-values in columns 1-4 are adjusted for estimation of the inflection point, which is estimated in column 5 with 95% confidence limits. Columns 6-7 show p-values testing the null of simple linearity against the alternative that 2 line segments are necessary for the regression of VA on CST. Column 8 gives the Pseudo score p-value for the null of 2 segments against the alternative of 3 segments. For further explanation, see text.

Columns 6 and 7 of Table 1 show Randomization and pseudo Score p-values testing the null hypothesis of a simple linear association against the alternative that 2 line segments are necessary to depict the relationship between CST and concurrent VALS, and testing the necessity of 3 line segments, at baseline and Months 1, 6, 12, 24, 36, 48, and 60. As suggested in Figure 1, 2-segment models are favored over 1-segment (linear) models for post-baseline months (perhaps except for Months 48 and 60, where the Randomization and pseudo Score tests disagree as to whether 1 or 2 segments are needed), and a 1-segment model suffices for baseline data. More segments may be necessary for Month 1, but not at other visits (column 8). The estimated slopes and intercepts that define the regression lines shown in Figure 1, which arise from the Randomization approach, are almost identical to those from the Segmented package approach (data not shown).

Discussion

With the possible exception of Months 48 and 60, when results are impacted by smaller sample sizes and a more complicated treatment history than at the earlier visits, the pseudo Score test and Randomization test agree that a 2-segment piecewise linear model fits post-baseline visit data significantly better than a linear model with a single slope parameter, suggesting that the CST-VALS relationship after treatment is not linear. Estimated inflection points are roughly constant over post-baseline time, ranging non-systematically from 217 microns to 256 microns using Reading Center data. There are strongly positive correlations to the left of the inflection points (where VALS increases with increasing CST) and strongly negative correlations to the right of the inflection points (where VALS decreases with increasing CST). This makes sense since, to the left of the inflection point, the retina is thinner than normal and the retinal tissue may be too thin to support good vision so that as the retinal thickness increases and approaches normal, the VALS improves (positive correlation) until the inflection point is reached; to the right of the inflection point, the retina is thicker and the thickened retina with associated fluid degrades VALS so that as the retinal thickness increases to the right of the inflection point, the VALS decreases (negative correlation) (Figure 1).

Our finding that, in SCORE2, the CST-VALS relationship is nonlinear is similar to what was found in eyes with neovascular age-related macular degeneration in the Comparison of Age-related Macular Degeneration Treatments Trials (CATT). At all follow-up time points in CATT, eyes with central retinal thickness (measured as foveal center retinal thickness by time domain OCT; Stratus; Carl Zeiss Meditic, Jena, Germany) between 120 and 212 microns had better visual acuity than eyes with foveal center thickness <120 microns and eyes with thickness >212 microns.³⁵ The lower optimal central retinal thickness values in CATT compared to those in SCORE2 is likely because time domain OCT was used in CATT while SD-OCT was used in SCORE2.³⁶ It should also be noted that CATT investigated patients treated with ranibizumab or bevacizumab for neovascular age-related macular degeneration while SCORE2 investigated patients treated initially with aflibercept or bevacizumab for macular edema associated with CRVO or HRVO. The CATT authors stated that “[a]s expected, abnormally thick retinas had decreased VAs” and postulated that geographic atrophy may have been one of the causes of retinal thinning and associated decreased VA.³⁵ In contrast, in SCORE2, retinal thinning may have been associated with decreased VALS due to ischemic retinal damage from retinal vein occlusion.

At baseline, prior to treatment in SCORE2, the 2-segment model is not supported by the data (P=0.56), and a simple linear relationship fits the data, with an overall negative correlation (r=−0.47, P<0.01; Table 1). This makes sense and is consistent with our model, since at baseline, almost all the eyes had a CST well above the estimated post-baseline inflection points.

Normative values for CST measured using the Heidelberg Spectralis SD-OCT in otherwise healthy eyes have been reported to be 275.2 + 24.2 microns for individuals aged 20-40 years, 269.4 + 22.1 microns for individuals aged 41-60 years, and 263.0 + 20.2 microns for individuals aged 61 and older;³⁷ this is consistent with findings from the current study, in which SCORE2 participants with a post-treatment CST between 217 and 256 microns had the best VALS. Of note, 79.1% of SCORE2 participants were aged 60 years or older, 14.9% were aged 50-<60 years, and only 6.1% were <50 years of age; the mean age of SCORE2 participants was 69 years.¹⁸

In SCORE2, the overall correlation assuming a linear relationship between CST and VALS is −0.20 at Month 1 and −0.16 at Month 12, and is a little higher at the longer-term follow-up visits (r=−0.51 at Month 24 and r=−0.29 at Month 60), but the detailed analyses presented in this report demonstrate that these overall usually weak correlations belie the strong correlations that are present in 2 different segments because, at each follow-up time point, the correlation to the left of the inflection point is strongly positive and the correlation in to the right of the inflection point is strongly negative, such that the correlations in the 2 segments partially cancel each other out. This may explain, at least in part, the usually poor to modest correlations between OCT-measured central retinal thickness and visual acuity reported in patients with diabetic macular edema^21-27 or macular edema associated with retinal vein occlusion.²⁸ Our results also indicate that, among patients treated with anti-vascular endothelial growth factor therapy for macular edema due to CRVO or HRVO, a retina which is too thin cannot support good visual acuity (this is supported by the positive correlation to the left of the inflection point), and that individuals with a CST close to the inflection point have the best visual acuity. It may be that the Month 1 data require more than two segments (Table 1, column 8), but this does not vitiate one of the central messages of this analysis, which is that, when it comes to the association between CST and VALS, a thinner retina is not always better.

Study limitations include SCORE2 attrition at later visits. As reported previously, 330 of the original 362 patients randomized into SCORE2 were enrolled in the SCORE2 Long Term Follow-up Study.²⁰ Of these, 36 (11%) are known to have died, and 150 (51%) of the remaining 294 completed a Month 60 visit. How the SCORE2 participants who did not complete long-term follow-up visits would have affected the observed relationship between VALS and CST if their data were available is unknown. Additionally, treatments were at investigator discretion after Month 12, which limits our ability to optimally compare the originally randomized treatment groups after Month 12, but treatment by investigator discretion is more likely to approximate “real world” outcomes. Further, SCORE2 included only patients treated with anti-VEGF therapy for macular edema due to CRVO or HRVO, so that the pattern of association, and the inflection point(s) if they exist, may be different for patients treated for macular edema due to other etiologies. Moreover, we demonstrate the nonlinear relationship only for CRVO and HRVO patients who have been treated with anti-VEGF. Our baseline data show that a linear relationship suffices to describe the CST-VALS relationship for patients with macular edema due to CRVO or HRVO before treatment.

In summary, in this SCORE2 analysis we found: 1) a nonlinear and modest to strong relationship between post-baseline CST and concurrent VALS (this nonlinear relationship may explain, at least in part, the range of the usually poor to modest overall correlations between OCT-measured central retinal thickness and visual acuity reported previously among patients with macular edema due to diabetic macular edema and retinal vein occlusion); 2) participants with a post-treatment CST close to 217 to 256 microns had the best VALS; and 3) a thinner retina is not always associated with better VALS.