Lifetime Outcomes of Anti–Vascular Endothelial Growth Factor Treatment for Neovascular Age-Related Macular Degeneration

Key Points

Question  What are lifetime outcomes of anti–vascular endothelial growth factor treatment for neovascular age-related macular degeneration?

Findings  In this multistate model using real-world data, an estimated 12% of the sample retained driving visual acuity and 15% reading visual acuity in at least 1 eye over their mean remaining lifetime of 11 years. More injections and younger age at baseline were associated with better outcomes.

Meaning  Good lifetime vision outcomes can be achieved with anti–vascular endothelial growth factor treatment in neovascular age-related macular degeneration with more injections and an earlier start, increasing the chances of good outcomes and highlighting the need to provide easy access to this treatment.

Importance  Neovascular age-related macular degeneration (nAMD), the largest single cause of irreversible severe vision loss in high-income countries, can now be treated with vascular endothelial growth factor (VEGF) inhibitors, but to our knowledge, no data on lifetime outcomes are available.

Objective  To determine visual acuity (VA) outcomes of anti-VEGF treatment for nAMD in both eyes for patients’ remaining lifetime.

Design, Setting, and Participants  Multistate modeling using real-world cohort data of 3192 patients with nAMD (>67 000 visits) treated in routine eye clinics in Australia, New Zealand, and Switzerland. Data were analyzed between 2007 and 2015.

Exposures  Intravitreal anti-VEGF treatment at the treating physician’s discretion and prospective data collection in standardized registry.

Main Outcomes and Measures  Visual acuity in both eyes over the remaining lifetime.

Results  For the mean remaining lifetime of 11 years, an estimated 12% (n = 371; 95% CI, 345-400) of the sample retained driving VA and an estimated 15% (n = 463; 95% CI, 434-495) reading VA in at least 1 eye. At that time, an estimated 82% of the sample (n = 2629; 95% CI, 2590-2660) had dropped out. Younger age at baseline and more injections during the first year of treatment were associated with better long-term outcomes.

Conclusions and Relevance  Anti-VEGF treatment was associated with preserved useful visual acuity in almost 20% of patients over their average remaining lifetime. More than 80% of patients will cease treatment over that time, having likely experienced a deterioration of vision beforehand. This is a remarkable outcome compared with outcomes without intervention, which lead to legal blindness within 3 years of disease onset in 80% of those affected. These findings underline the public health necessity of providing anti-VEGF treatment to persons in need.

The late-stage neovascular form of age-related macular degeneration (nAMD) is the largest single cause of irreversible, severe vision loss in all high-income countries, causing estimated direct costs of more than €50 billion (US $59 billion) in the EU alone each year.^1,2 Current treatment options that preserve sight and lead to a considerable improvement in some patients are based on the inhibition of vascular endothelial growth factor (anti-VEGF) and are delivered as regular intravitreal injections.³ In contrast, nonintervention outcomes lead to legal blindness within 3 years following onset in 80% of patients affected.⁴ Because anti-VEGF treatment does not cure nAMD but only controls it, treatment is for the long term and is often required until either the eye worsens to levels where treatment is no longer considered beneficial or the patient dies.

However, to date, only a few reports on real-world outcomes beyond 5 years are available, and attrition is high, as is inevitable in long-term observational studies in elderly populations.^5,6 Thus, we do not know whether and for how long patients retain useful visual acuity (VA) that allows them to continue to undertake activities that were often lost prior to the advent of anti-VEGF treatment, such as reading and driving. Considering the considerable societal effects of both nAMD and its treatment in terms of costs, a better understanding of lifetime outcomes of anti-VEGF treatment for nAMD is required. Long-term treatment outcomes or prognosis can be estimated using multistate models (MSMs), which capture transition between different health states over time conditional on specific transition probabilities based on real-world data.⁷ In this study, we used MSMs as well as data from patients with long-term, ongoing anti-VEGF treatment for nAMD in routine medical practice to evaluate outcomes for this sample over their average remaining lifetime.

This was a retrospective analysis of an observational study of patients who had commenced intravitreal therapy for nAMD in routine clinical practice and had been prospectively tracked in the Fight Retinal Blindness (FRB) database. Treatment decisions and visit schedules were determined by the treating physician in consultation with the patient. Both visits and treatment were at maximum monthly. The details of the FRB database have been previously published.⁸ In brief, the FRB registry collects data from each clinical visit. Institutional ethics approval was obtained from the human research ethics committees of the University of Sydney, the Royal Victorian Eye and Ear Hospital, the Royal Australian and New Zealand College of Ophthalmologists, and the Cantonal Ethics Committee Zurich, and patients provided written consent. No compensation or other incentives were provided for participation. The study adhered to the tenets of the Declaration of Helsinki.

All patients resided in Australia, New Zealand, or Switzerland; were aged 55 years or older with a diagnosis of nAMD and received treatment with anti-VEGF between 2007 and 2015; and had VA data for both eyes and at least 1 follow-up. Patients were excluded if data were missing on which eye was treated. Patients were not excluded based on ocular comorbidities or if they underwent cataract surgery during follow-up. Best-corrected VA was tested using Early Treatment Diabetic Retinopathy Study charts or logMAR charts. Treatment was with anti-VEGF agents only; no patients received intravitreal corticosteroids.

Multistate modeling requires discrete health states through which patients transition with a certain probability. Because the assessed outcome is VA, we considered discrete vision states that have been described in more detail elsewhere.⁹ In brief, 3 categories of vision impairment (VI) were calculated for both eyes, individually: (1) no VI (VA ≥20/40); (2) mild VI (<20/40 to 20/60); and (3) moderate to severe VI (VA <20/60). These cutoffs were chosen because 20/40 is the cutoff for driving VA in most countries and 20/60 is a cutoff for comfortable reading VA (no need for magnifying aids).^10,11 With 3 categories of vision impairment in each eye, a combination of 6 different vision states were determined based on the 3 categories of VI in each eye (eTable in the Supplement). Treatment cessation, dropout, and death were all classified as dropout because the data did not allow for a differentiation. Patients were classified dropout if no further visits were documented for at least 6 months. Any documented visits, irrespective of the interval between them, were included in the respective vision state in the model.

Descriptive statistical analysis methods were used to characterize the sample. These included means with standard deviations (SD) and medians with interquartile ranges (IQRs) for continuous variables and numbers with percentages for categorical variables.

We used multistate Markov models to describe the process in which an individual transitions through a series of vision states in continuous time while receiving anti-VEGF treatment for nAMD. A Markov model is a general MSM in which a system switches between different states assuming a certain transition probability. Multistate Markov models are commonly used in studies of chronic diseases, in which patients are assumed to pass through a series of discrete disease stages with 1 final state, either death or the terminal stage of a disease from which no exacerbation is possible, that is irreversible and referred to as an “absorbing” state.^12,13 This approach works very well for censored data¹⁴ and has been implemented in the software solution R (R Statistical Foundation; add-on package msm).¹⁵

With the addition of a dropout state, ie, treatment cessation for any reason including mortality, as an absorbing state, our data reflected a 7-state system: 6 vision states as well as 1 absorbing health state in which patients were no longer treated. Transitions between any of the 6 vision states are possible; however, an improvement for example from vision state 6 to vision state 1 is unlikely (see eFigure 1 in the Supplement for possible transitions). Our data were entered into a multistate hidden Markov model, with monthly transitions between health states. Patients who were extended beyond 1 month were classified as remaining in the vision state until they were assessed the next time, ie, had their VA recorded again, at which point they could either remain (ie, no change) or transition into another vision state (ie, improve, deteriorate, or dropout). At any time, any individual is in a discrete vision state or in the absorbing health state. Transition between and time within each of the vision states are governed by the according transition probabilities.

The msm functions in the R package of the same name, version 1.6.6, were then used to estimate a transition probability matrix and the remaining mean life expectancy of patients based on the Australian Life Tables (Australian Bureau of Statistics at https://www.abs.gov.au/ausstats/). Because age at baseline and the number of injections received during the first year are the most important factors associated with treatment outcomes besides baseline VA,¹⁶ MSMs were adjusted for the most populous age groups (aged 70-79 years and aged 80-89 years) and by number of injections during the first year of treatment (≤5, 6-7, or ≥8 injections). Only the initial number of injections was used to characterize transition rates, whereas subsequent variations of the injection frequency were not included as a predictor in our model. The calculation of life expectancy was also adjusted for age at baseline and sex. For each transition probability (over 5 and 10 years), 95% confidence intervals (CIs) were calculated using a nonparametric bootstrap approach, with 500 resamples, as implemented in the msm package.¹⁵

The goodness of model fit was assessed by comparing expected and observed prevalence (or occupancy) in all health states per month for the first 5 years for which we had real-life data. This accurately describes how well the model captures the assessed process over time and is an appropriate approach for assessing goodness of fit of hidden Markov models.¹³ All analyses were undertaken using the R software for statistical computing, version 3.5.2.

We obtained complete data sets of 3192 individual patients with a total of 67 700 visits (Table 1). Thirty-seven percent were male and 63% female, and most were 80 years or older at baseline (Table 1). Thirty-nine percent of the sample were treated in both eyes. On average, patients were followed up for just under 3 years (mean [SD], 34 [37] months), and mean (SD) VA was similar at last follow-up compared with baseline (60 [22] and 61 [19] letters). Patients received a mean (SD) of 18 (17) injections during follow-up ( Table 1).

Estimated transition probabilities and 95% CIs from baseline to 5 and 10 years are listed in Table 2. Using vision state 1 as an example, patients with good VA in both eyes (vision state 1, VA ≥20/40) at baseline had a 13% chance at 5 years and a 5% chance at 10 years to remain at that level of vision in both eyes. However, they also had a 17% chance of losing vision in one (transition to vision states 1 and 3) and a 13% chance of losing vision in both eyes (transition to vision state 4-6) at 5 years and 8% and 6% chance at 10 years, respectively. Irrespective of baseline VA, patients had a 58% to 66% chance at 5 years and a 84% to 87% chance at 10 years of transitioning into dropout (Table 2).

Based on the estimated transition probabilities, we calculated treatment outcomes at 5 and 10 years and for the remaining life expectancy for our sample (Table 3). Average life expectancy was 11 years (mean, 10.9; median, 9.7; range, 2.1-31.3 years). After 5 years of anti-VEGF treatment, 26% of the sample (n = 841; vision states 1-3; 95% CI, 800-884) retained driving VA in at least 1 eye. Similarly, a total of 32% of the sample (n = 1020; vision states 1-5; 95% CI, 978-1068) retained reading VA in at least 1 eye after 5 years of treatment (Table 3) while 62% of the sample (n = 1976; 95% CI, 1925-2022) had dropped out. Over the remaining life expectancy, 12% of the sample (n = 371, vision states 1-3; 95% CI, 345-400) retained driving VA, and 15% (n = 463, vision states 1-5; 95% CI, 434-495) retained reading VA in at least 1 eye. At that time, 82% of the sample (n = 2629; 95% CI, 2590-2660) had dropped out (Table 3).

Younger age at baseline as well as a higher number of injections during the first year were associated with better functional outcomes at both 5 and 10 years (Figure). In all modeled scenarios, both for driving and reading VA and both after 5 and 10 years, VA outcomes improved steadily from 5 or fewer injections to 6 to 7 injections to 8 or more injections given during the first year of treatment. Similarly, in all modeled scenarios, persons aged 70 to 79 years at baseline compared with those aged 80 to 89 years had better outcomes. Patients with 8 or more injections were treated for longer before they dropped out (mean, 3.3 vs 2.4 years; median, 2.8 vs 1.4 years; first quartile, 1.8 vs 0.4 years; third quartile, 4.6 vs 4.2 years). Patients with good baseline vision, ie, driving and reading VA (vision state 1-3), had the highest probability to retain good visual acuity at both 5 and 10 years (Table 2) as well as over their remaining lifetime (Table 3) irrespective of their baseline age.

The observed and expected prevalence rates for each vision state and the absorbing health state are depicted in eFigure 2 in the Supplement. Vision states 4 and 5 demonstrate the best fit (eFigure 2 in the Supplement). Those who ceased treatment or dropped out for any other reason such as loss to follow-up or death are somewhat underestimated from month 40 onwards (panel 7 in eFigure 2 in the Supplement).

With anti-VEGF treatment of nAMD in routine medical practice, a large proportion of patients retained useful VA in at least 1 eye for 5 years and a smaller proportion over their estimated remaining lifetime. An increasing proportion of patients will cease treatment over time, with more than 80% not being treated any longer toward the end of their remaining life expectancy. These findings emphasize that ongoing anti-VEGF treatment is associated with better outcomes and prevention of severe vision loss in a considerable proportion of patients in the long term. It also highlights that a large proportion of patients cease treatment for various reasons, and more work is required to lower this proportion.

Several studies have demonstrated a reduction in blindness registrations owing to nAMD of up to 50% since the introduction of anti-VEGF treatment,^17-19 but no study to our knowledge has assessed whether patients retain useful vision instead of just avoiding legal blindness. The natural history of nAMD is devastating, with almost 80% of patients (77.6%) becoming blind (VA ≤20/200) by 3 years.⁴ Based on our modeled lifetime outcomes, the effect of anti-VEGF treatment is immense, with a considerable proportion retaining useful VA in at least 1 eye rather than just avoiding legal blindness. However, a large proportion of patients will cease treatment (ie, drop out), which is reflected in virtually all real-world case series in the literature, with treatment cessation in the first year of up to 50%.²⁰

Multistate models are commonly used to assess disability or chronic disease progression with discrete health states.^7,13,21 Using MSMs to assess and extrapolate VA outcomes has not been done before to our knowledge, most likely because VA is a continuous variable and not commonly thought of as a vision or health state. However, given the cutoffs we chose in this study, it is not only possible but sensible to assess the proportion of patients who retain VA useful for driving and/or reading, ie, avoid visual disability.

Our model may slightly overestimate prevalence in several vision states and underestimate transition into the dropout state. Because the dropout state includes persons who discontinued treatment with good vision owing to disease inactivity as well as persons who continued treatment at another clinician with potentially good visual outcomes as well, albeit only small numbers in both groups, the dropout state includes persons with good VA outcomes. Thus, the observed prevalence in the dropout state is higher than the expected, ie, modeled, and higher than the real prevalence. In addition, this group of elderly patients has a high and increasing background mortality over the last 10 years of their life, which was impossible to account for in a separate health state owing to the study design.

Strengths of this study include the setting in routine medical practice, capturing real-life treatment patterns and outcomes. Our treatment outcomes are in keeping with published literature of real-world outcomes and we considered treatment and outcomes for both eyes because this reflects clinical reality. In previous work, we have demonstrated that independent of the better eye, the worse eye significantly affects a person’s quality of life and utilities, and the effect of treatment on VA and thus patients’ quality of life is independent of whether the better or worse eye is treated.^9,22,23

The MSM used in this work has the advantage that it analyzes the driving forces behind the transitions between vision states, accounting for both risk factors and the lengths of the time spans between transitions. Of note, the assumed structure of our transition probabilities is related to a 2019 modeling approach by Grand et al,²⁴ who analyzed the effects of risk factors and time spans by fitting a Cox-type MSM with a constant baseline hazard to analyze visual acuity (measured on a continuous scale) within a joint modeling framework. Unlike this model, the MSM for vision states considered here is not based on eye-level measurements but on person-level measurements because this is the most appropriate level when assessing patients’ everyday functioning.

Limitations include the comparatively small sample size for such a modeling study, the relatively short follow-up in comparison with the remaining lifetime modeled, the somewhat inhomogeneous data characteristic for registries and case series, and the attrition of patients over time. However, our sample size is no smaller than samples used in other MSM studies for outcome prediction,^7,13,21 and the uncertainty in our estimates is well reflected in the widths of the bootstrap confidence intervals. Also, the MSMs considered here explicitly account for attrition over time because we included a separate state to account for death and/or dropout. While MSM is an excellent means to extrapolate transition between health states in continuous time, predicting future outcomes with as much accuracy as statistically possible, we emphasize that the specification of our model is not without limitations. First, owing to the sampling design and the characteristics of the data set, only 2 covariates (age at baseline and number of injections) of potentially many variables affecting the transition probabilities could be analyzed. Second, again owing to the sampling design, we had to collapse both “death” and “dropout for other reasons” into 1 state, which implied that our model was not able to separate the (potentially different) associations between these events and the various vision states. Still, despite these limitations, our probability estimates and conclusions are plausible and appear to be well supported by the data at hand. Third, including all FRB participants irrespective of ocular comorbidities or cataract surgery during follow-up affected our ability to attribute VA gains or losses to anti-VEGF treatment. However, it increases external validity because it reflects actual patient cohorts treated in real life. Fourth, the dropout state likely contains patients with good VA who discontinued treatment owing to disease inactivity, which should be considered when interpreting the data. Finally we note that generalizing results from a setting in the Australian, New Zealand, or Swiss health care system requires caution. Similarly, generalizing results to samples with different sex or age distributions, injection frequencies, or a different distribution of baseline VA requires caution.

In conclusion, anti-VEGF treatment preserves useful VA in almost 20% of patients for their average remaining life span. However, a large proportion of patients will cease treatment over that time, of whom many may have experienced a deterioration of VA. Given the almost uniformly rapid progressive vision loss associated with the natural history of nAMD, this is a remarkable outcome for any chronic disease, underlining the public health necessity of providing anti-VEGF treatment to persons in need as early as possible. However, more can be done to understand how to increase the proportion of those maintaining good VA over their remaining life span, so that more people are able to age well.

Corresponding Author: Robert P. Finger, MD, PhD, Department of Ophthalmology, University of Bonn, Ernst-Abbe-Str 2, 53127 Bonn, Germany (robert.finger@ukbonn.de).

Accepted for Publication: August 22, 2020.

Published Online: October 15, 2020. doi:10.1001/jamaophthalmol.2020.3989

Author Contributions: Drs Finger and Schmid had full access to all of the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis.

Concept and design: Finger, Schmid, Guymer.

Acquisition, analysis, or interpretation of data: All authors.

Drafting of the manuscript: Finger.

Critical revision of the manuscript for important intellectual content: All authors.

Statistical analysis: Puth, Schmid.

Obtained funding: Finger, Barthelmes, Guymer.

Administrative, technical, or material support: Finger, Barthelmes, Guymer, Gillies.

Supervision: Finger, Schmid, Guymer.

Conflict of Interest Disclosures: Dr Schmid reported personal fees from Pixium Vision outside the submitted work. Dr Barthelmes reported grants from Novartis and Bayer during the conduct of the study and personal fees from Alcon outside the submitted work; in addition, Dr Barthelmes had a patent to Fight Retinal Blindness software issued. Dr Guymer reported personal fees from Roche outside the submitted work. Dr Gillies reported grants and personal fees from Novartis and Bayer during the conduct of the study and outside the submitted work. Dr Finger reported personal fees from Novartis, Bayer, Roche/Genentech, Allergan, Alimera, Santhera, Opthea, Inositec, and Ellex outside the submitted work. No other disclosures were reported.

Funding/Support: This work was in part supported by the German Scholars Organization/Else Kröner Fresenius Stiftung (GSO/EKFS-16). The Fight Retinal Blindness Project was supported by a grant from the Royal Australian and New Zealand College of Ophthalmologists Eye Foundation (2007-2009); a grant from the National Health and Medical Research Council, Australia (NHMRC 2010-2012); and a grant from the Macular Disease Foundation, Australia. Dr Gillies is a Sydney Medical Foundation Fellow and is supported by an National Health and Medical Research Council, Australia practitioner fellowship. Dr Barthelmes was supported by Walter Gertrud Siegenthaler Foundation, Zurich, Switzerland, and the Swiss National Foundation. Dr Guymer was supported by National Health and Medical Research Council of Australia fellowship grant GNT1103013. Centre for Eye Research receives operational infrastructure support from the Victorian government.

Role of the Funder/Sponsor: The funding sources had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.