Efficacy and Adverse Events of Anti-VEGF Agents in AMD
This systematic review followed the PRISMA statement. Ethics approval was not required for this network meta-analysis because only published data were included.
We included randomised controlled trials comparing aflibercept, bevacizumab or ranibizumab against placebo or in a head-to-head fashion. Studies had to include 1-year follow-up data of visual acuity and serious side effects.
To identify randomised controlled trials, we applied an iterative approach combining electronic database and hand searches. The search strategy was designed in collaboration with an information specialist. We searched papers in Medline, Premedline, EMBASE, SCOPUS and the Cochrane Library. Searches were performed separately for papers examining the efficacy and serious side effects of aflibercept, bevacizumab and ranibizumab in patients with AMD. The search was last updated in June 2013. From papers qualifying for inclusion we examined related articles using the related article function available on the PubMed Medline interface. Moreover, we checked for articles citing three seminal papers using the Science Citation Index Database. Examination of reference lists of included studies and review articles complemented our searches. The Medline search strategy is available in the online supplementary appendix.
In duplicate, we extracted salient methodological features (description of generation of random sequence and concealment of random allocation, blinding of patients and caregivers, whether the analysis was based on the intention to treat principle and the proportion of patient lost during follow-up), patient characteristics (patients' age, previous laser treatment of AMD, type and size of membranes), and treatments and dosages of each paper. Discordant extractions were discussed. This was the case in 18 instances and was due to an over-reading.
Efficacy. The most consistently reported efficacy outcome was increase in letters gained. We chose this outcome for this network meta-analysis.
Side Effects. We initially extracted all reported serious side effects from each included study. In a second step, we selected those that were at least reported three times and were considered as serious. The analysis thus considered the following five outcomes: vascular death, any death, stroke, myocardial infarction and transient ischaemic attack. In a separate analysis, we looked at systemic thrombotic events.
Details about this statistical approach and its application have been published elsewhere.
Efficacy Analysis. Results from intention to treat analysis were considered wherever possible.
We imputed missing SDs of mean changes for each treatment using the largest SD reported in the set of included studies for this outcome. This was necessary in four cases. For each participant, we simulated the outcome by sampling from a normal distribution with mean and SD of the outcome in a specific treatment arm as described in the study report. Due to chance, the mean and SD parameters could be different from the original values. We corrected these differences using a linear transformation. We generated such a dataset for all the treatment arms. This approach led to the same likelihood functions as that from the original data. To this new dataset, a linear regression model was fitted. Drug and dosage, creating a unique code for each treatment, were entered as covariates. An indicator variate for each study was entered to the model to preserve randomisation within each trial. This variate adjusted for all the possible differences (patients and design) between studies.
Side Effects Analysis. For each of the serious side effects and each of the treatment arms, the number of events was added up and then divided by the total number of patients in the corresponding treatment arm. Thus, the events per patient in each of the trials given a specific treatment and dosage were determined. The total score of serious side effects was calculated by adding up these estimates of all serious side effects. For the sum of serious side effects outcomes, a linear regression analysis was performed with drug, dosage and an indicator for the studies as covariates. As a substitute for the inverse of the variance, we weighted the analysis with the total number of patients in each treatment arm.
Trade-off Analysis. For efficacy, we looked at number of letters gained. This efficacy parameter was each plotted against serious side effects and thrombotic events from the network meta-analysis. Results were calculated as percentage differences against placebo. Analyses were performed with the Stata SE V.11.2 software package (Copyright 1996–2010 StataCorp LP, 4905 Lakeway Drive, College Station, TX 77845, USA).
Methods
This systematic review followed the PRISMA statement. Ethics approval was not required for this network meta-analysis because only published data were included.
Eligibility Criteria for Considering Studies for This Review
We included randomised controlled trials comparing aflibercept, bevacizumab or ranibizumab against placebo or in a head-to-head fashion. Studies had to include 1-year follow-up data of visual acuity and serious side effects.
Search Methods for Identifying Studies
To identify randomised controlled trials, we applied an iterative approach combining electronic database and hand searches. The search strategy was designed in collaboration with an information specialist. We searched papers in Medline, Premedline, EMBASE, SCOPUS and the Cochrane Library. Searches were performed separately for papers examining the efficacy and serious side effects of aflibercept, bevacizumab and ranibizumab in patients with AMD. The search was last updated in June 2013. From papers qualifying for inclusion we examined related articles using the related article function available on the PubMed Medline interface. Moreover, we checked for articles citing three seminal papers using the Science Citation Index Database. Examination of reference lists of included studies and review articles complemented our searches. The Medline search strategy is available in the online supplementary appendix.
Data Collection and Risk of Bias Assessment
In duplicate, we extracted salient methodological features (description of generation of random sequence and concealment of random allocation, blinding of patients and caregivers, whether the analysis was based on the intention to treat principle and the proportion of patient lost during follow-up), patient characteristics (patients' age, previous laser treatment of AMD, type and size of membranes), and treatments and dosages of each paper. Discordant extractions were discussed. This was the case in 18 instances and was due to an over-reading.
Outcome Measures
Efficacy. The most consistently reported efficacy outcome was increase in letters gained. We chose this outcome for this network meta-analysis.
Side Effects. We initially extracted all reported serious side effects from each included study. In a second step, we selected those that were at least reported three times and were considered as serious. The analysis thus considered the following five outcomes: vascular death, any death, stroke, myocardial infarction and transient ischaemic attack. In a separate analysis, we looked at systemic thrombotic events.
Data Synthesis and Analysis
Details about this statistical approach and its application have been published elsewhere.
Efficacy Analysis. Results from intention to treat analysis were considered wherever possible.
We imputed missing SDs of mean changes for each treatment using the largest SD reported in the set of included studies for this outcome. This was necessary in four cases. For each participant, we simulated the outcome by sampling from a normal distribution with mean and SD of the outcome in a specific treatment arm as described in the study report. Due to chance, the mean and SD parameters could be different from the original values. We corrected these differences using a linear transformation. We generated such a dataset for all the treatment arms. This approach led to the same likelihood functions as that from the original data. To this new dataset, a linear regression model was fitted. Drug and dosage, creating a unique code for each treatment, were entered as covariates. An indicator variate for each study was entered to the model to preserve randomisation within each trial. This variate adjusted for all the possible differences (patients and design) between studies.
Side Effects Analysis. For each of the serious side effects and each of the treatment arms, the number of events was added up and then divided by the total number of patients in the corresponding treatment arm. Thus, the events per patient in each of the trials given a specific treatment and dosage were determined. The total score of serious side effects was calculated by adding up these estimates of all serious side effects. For the sum of serious side effects outcomes, a linear regression analysis was performed with drug, dosage and an indicator for the studies as covariates. As a substitute for the inverse of the variance, we weighted the analysis with the total number of patients in each treatment arm.
Trade-off Analysis. For efficacy, we looked at number of letters gained. This efficacy parameter was each plotted against serious side effects and thrombotic events from the network meta-analysis. Results were calculated as percentage differences against placebo. Analyses were performed with the Stata SE V.11.2 software package (Copyright 1996–2010 StataCorp LP, 4905 Lakeway Drive, College Station, TX 77845, USA).
SHARE
