Eligibility criteria

This systematic review will include data from randomized controlled trials, cohort studies, and case series that meet all of the following criteria: 1) involve patients with ERM undergoing surgery, 2) assess OCT biomarkers measured before surgery (ideally within one year prior to surgery) to determine their predictive value for treatment response, either individually or as part of a multivariable clinical prediction model.

Criteria for exclusion are: 1) Studies including secondary ERMs associated with other ocular conditions (e.g., diabetic macular edema, retinal vein occlusion, rhegmatogenous retinal detachment, uveitis, or fibrovascular proliferation); 2) Studies including eyes with full-thickness macular hole; 3) Studies including post-pars plana vitrectomy ERMs; 4) studies including eyes with full-thickness macular hole (partial-thickness tractional holes, epiretinal membrane foveoschisis, and lamellar macular holes will be included) or with advanced macular co-pathology affecting vision; 5) studies with less than 10 patients with ERM; 6) studies describing techniques or guidelines; 7) studies in which surgical parameters are the primary measures of outcome; 8) studies with no stratified analysis for patients with ERM; 9) non-original research/review.

We are interested in the following outcomes at 6 months, 12 months, or 24 months after surgery: [9]:

• Final postoperative best-corrected visual acuity (BCVA) and mean change in BCVA, measured using a standardized chart (e.g., Logarithm of the Minimum Angle of Resolution (logMAR), Snellen) at a starting distance of 4 m.
• Proportion of patients with a gain of>= 0.2 logMAR in BCVA or uncorrected visual acuity (UCVA) in the study eye.
• Proportion of patients with a loss of>=0.2 logMAR in BCVA or UCVA in the study eye.
• Mean change in quality of life (QoL), measured using a validated questionnaire.
• Any adverse events (AEs) identified during follow-up.

A threshold of 0.2 logMAR (10 Early Treatment Diabetic Retinopathy Study (ETDRS) letters or 2 Snellen lines) will be selected because this value is recognized in ophthalmology research [9] and has been adopted in clinical trial of retinal diseases [19]. We will search without restrictions on language or date of publication.

Search strategy

With support from an experienced research librarian, we will search Ovid MEDLINE, EMBASE, CENTRAL and Web of Science using a combination of keywords and Medical Subject Headings (MeSH) terms. We will systematically search from inception. We will supplement our search with backward and forward citation searching using citationchaser (https://estech.shinyapps.io/citationchaser/) and search the reference lists of systematic reviews identified [9,16,20]. To ensure inclusion of all relevant studies, we will contact clinical experts in the field.

Screening

After training and calibration exercises, reviewers will work independently and in duplicate to screen titles and abstracts of search records and, subsequently, full texts to determine eligibility. All disagreements will be resolved by discussion and, when necessary, by a third reviewer. We will use Covidence (Melbourne, Australia, https://www.covidence.org), an online systematic review software for screening titles and abstracts and full-text articles. A PRISMA Flow Diagram will provide a visual overview of the records identified, included and excluded, and the reasons for exclusion [21].

Data extraction

Two reviewers, working independently and in duplicate, will collect data using a standardized pilot-tested collection form designed in Microsoft Excel (Version 16.52). Discrepancies in data collection will be resolved by discussion and disagreements will be resolved with a third reviewer. We will contact the corresponding author or the sponsoring pharmaceutical company for clarification in case of missing data.

For each included study, we will extract study characteristics, patient characteristics, surgical techniques, biomarker(s) examined and their definitions, whether an image reading center was employed to identify biomarkers, the association between baseline biomarker(s) and visual outcomes of interest (we listed the detailed items in Table 1). We listed the potential biomarkers and definitions in Table 2. Other structural OCT biomarkers not listed in Table 2 will also be eligible for inclusion if reported.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Data extraction items.

https://doi.org/10.1371/journal.pone.0338330.t001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. The list of potential baseline structural OCT biomarkers and definitions.

https://doi.org/10.1371/journal.pone.0338330.t002

We will extract VA, QoL, and AEs at 6 months (±3 months), 12 months (±3 months) and 24 months (±3 months) after treatment initiation, type of analysis conducted, whether the analysis is univariate or multivariate. If multivariate, we will extract the variables controlled for in the analysis and whether the study follows the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD) guidelines [22]. We will extract whether bilateral eyes are allowed, and if so, how they are handled in the analysis.

Risk of bias assessment

We will use the Quality in Prognosis Studies (QUIPS) tool to assess risk of bias for prognostic factor studies [23]. For studies that develop or validate multivariable clinical prediction models, we will apply the Prediction model Risk Of Bias ASsessment tool (PROBAST) [24]. QUIPS tool assesses risk of bias across six domains: participation, attrition, prognostic factor measurement, confounding measurement and account, outcome measurement, and statistical analysis and reporting. There are four domains (participants, predictors, outcome, and analysis), and total of 20 signaling questions in PROBAST. Any disagreements will be resolved through discussion. For the risk of study confounding to be rated as low, we will pre-specify that the study must report a multivariable model that includes, at a minimum, age [20], patients’ baseline VA [16] and baseline retinal thickness [25]. The overall risk of bias will be deemed low if all domains are rated as low risk, and high if one or more domains are rated as high risk. Otherwise, the study will be rated as having a moderate risk of bias.

Data analysis

We will conduct meta-analyses using the restricted maximum likelihood random-effects model for all prognostic factors reported by two or more studies. VA is frequently reported using the ETDRS letter score. We will convert values reported in Logarithm of the logMAR, Snellen chart scores or any other validated scale to approximate ETDRS scores using the method proposed by Gregori et al [26]. If multiple scales measured the same outcome, we used linear transformation to the most used scale [27]. For dichotomous outcomes, we will convert relative effects into absolute effects to facilitate interpretation and magnitude of effect. We will calculate absolute effects using pooled RRs applying the median risk in control groups of studies that report 1 or more events as the baseline risk. If both eyes from the same patient are included in a study and analyzed as independent units without appropriate adjustment, we will consider this in the risk of bias assessment and explore the potential impact in sensitivity analyses. Where possible, we will extract effect estimates that account for within-patient correlation; otherwise, we may include only one eye per patient (e.g., randomly selected or as reported by study authors) to avoid double-counting.

If the same OCT biomarker is labeled differently across studies, we will extract each study’s definition and group those that appear sufficiently similar. To minimize bias, the definitions will be shared with three ophthalmologists blinded to study results, and they will confirm or revise our grouping decisions. When consensus cannot be reached, or definitions differ substantially, results will be summarized narratively. All analyses will be performed in R (Version 4.3.1) using the meta and metafor packages [28,29]. When quantitative analysis is not possible, we will present results narratively using the SWiM (Synthesis Without Meta-analysis) guidance [30]

If two or more studies are available for any subgroup, the following subgroup analyses will be conducted:

1. i) High vs. low risk of bias, hypothesizing that studies with higher risk of bias may overestimate the strength of associations between baseline OCT biomarkers and visual outcomes.
2. ii) Presence of a reading center, hypothesizing that studies not using a reading center may overestimate the associations [31,32].
3. iii) Lens status and surgery type (pars plana vitrectomy alone vs. combined phacoemulsification and pars plana vitrectomy), hypothesizing that lens removal independently improves VA, and phakic eyes tend to develop or progress cataract after vitrectomy.
4. iv) Internal limiting membrane peel vs. no peel, hypothesizing that internal limiting membrane peeling may reduce recurrence.
5. v) Govetto stage/ presence of ectopic inner foveal layers, hypothesizing that greater disease indicates chronic traction and neuroglial remodeling, which may limit postoperative visual recovery.
6. vi) Partial-thickness macular defects: presence vs. absence of partial thickness holes (partial-thickness tractional holes, epiretinal membrane foveoschisis, and lamellar macular holes). We hypothesize that eyes with these conditions will show less visual improvement after ERM surgery compared with eyes without such defects.
7. vii) Degree of adjustment for critical predictors of outcome, hypothesizing that models that adequately adjust for key predictors may yield more reliable effect estimates.

The credibility of the statistically significant subgroup effects will be assessed using the Instrument to assess the Credibility of Effect Modification Analyses (ICEMAN) [33]. We will only present stratified results if moderate to high credibility subgroup effects are found.

Publication bias will be assessed by visually inspecting the funnel plots for outcomes reported in ten or more studies. If less than ten studies are included for an outcome, publication bias will be assessed by evaluating the quality of the search and measures taken to obtain all possible evidence to inform the outcome of interest [34].

Certainty of evidence

The certainty of evidence for this systematic review will be evaluated using the Grading of Recommendations, Assessment, Development, and Evaluation (GRADE) approach for prognosis [35,36], based on several key factors: methodological limitations, indirectness, imprecision, inconsistency, and potential publication bias [36]. The ratings will be determined using a minimal clinically important difference (MCID) of 15 ETDRS letters [37,38]. For dichotomous outcomes, we will use an absolute risk difference of 10% as the MCID.

The GRADE ratings will provide an overall confidence level in the evidence. A “High” rating will indicate that the true effect is very likely close to the estimated effect. A “Moderate” rating will suggest that the true effect is probably similar to the reported estimate, while a “Low” rating will imply that the true effect may be substantially different. A “Very Low” rating will reflect a high degree of uncertainty, limiting the ability to draw firm conclusions. These ratings will be used to determine whether each biomarker is associated with improved, stable, or worsened VA. We will use standardized language that incorporates certainty of evidence when communicating our results (i.e., high certainty evidence presented with declarative statements, moderate certainty evidence with ‘likely’, low certainty evidence with ‘may’, and very low indicated by ‘very uncertain’) [39].