Types of studies

We included reports of prospective and retrospective longitudinal studies that investigated whether the prognostic factor, protease activity, was associated with healing of VLUs, as such studies can produce valid odds or hazard ratios of prognostic factors. We did not include studies that solely examined the validation of prediction models, but we did include prediction model development studies when they also reported associations of individual prognostic factors (protease activity) with the outcome under study (healing). We included randomised controlled trials (RCTs) analysed as cohort studies, provided interventions were taken into account in the analysis. We included case‐control studies only if there were no relevant eligible cohort studies for specific proteases. We included case‐control studies because the design can provide valid estimates of associations between a prognostic factor and outcomes provided that the follow‐up time is not too long. However, we acknowledge that case‐control studies often suffer from other deficiencies (Altman 2001): these were assessed in our risk of bias assessment. We did not include cross‐sectional studies or case reports because the association under study is inherently longitudinal. We included studies with any follow‐up period, given the prognostic nature of our objective.

We planned to consider evidence separately within the different phases of prognostic factor investigation: phase 1 (exploratory), and phase 2 (confirmatory) studies, which provide different levels of evidence (Hayden 2008; Riley 2013). Exploratory studies identify associations of many potential prognostic factors and outcomes. These studies measure associations between single factors and the outcome ('univariable' or 'univariate' associations), and provide the least conclusive information regarding the independence of a variable as a valid prognostic factor. Confirmatory studies aim to measure the independent effect of a prognostic factor on the relevant outcome while controlling for other factors. We planned to include all eligible studies in the review and, if possible, planned to restrict the results summary and any meta‐analyses to results from multivariable analyses. Where this was not possible, we included the results of univariable analyses, taking into account the phase of investigation in both the risk of bias assessment and GRADE rating on certainty of the summary estimates.

Targeted population

We included studies of people with a VLU, who were managed in any care setting and receiving any type of treatment. We expected the method of diagnosis of venous ulceration to vary, so accepted definitions as used in the included studies.

We included studies in people with VLUs alongside people with other types of wounds (e.g. arterial ulcers, pressure ulcers, diabetic foot ulcers) provided the results for people with venous ulcers were presented separately, or if most participants (at least 75%) had leg ulcers of venous aetiology. Where wounds were described only as leg ulcers without information as to aetiology, we assumed they were venous in origin because this is the most common type of leg ulcer.

We included studies that involved participants at any stage in their treatment pathway, and we recorded, where available, baseline data on the time since diagnosis and treatments previously given, together with ongoing treatments. We included studies that involved participants with any infection status at baseline and recorded any available data on this.

We excluded studies conducted solely in vitro and animal studies.

Types of prognostic factor

The prognostic factor was protease activity. Any type of protease or combination of proteases was eligible, either from the host or from bacteria, including:

• MMPs: gelatinases (MMP‐2 and MMP‐9); collagenases (MMP‐1, MMP‐8 and MMP‐13); stromelysins (MMP‐3, MMP‐10 and MMP‐11); metalloelastase (MMP‐12); matrilysins (MMP‐7 and MMP‐26); membrane‐type MMPs (MMP‐14, MMP‐15, MMP‐16, MMP‐17, MMP‐24 and MMP‐25); and other MMPs (MMP‐19, MMP‐20, MMP‐23 and MMP‐28);

• serine proteases: mast cell tryptases and chymases; human nucleophil elastase; cathepsin G; plasmin; uPA; t‐PA;

• proteases combined with other biological species: ADAM; ADAMTS.

Where possible, we combined these three major categories in the analyses, taking into consideration cut‐off points used (see 'Data synthesis' section below).

We planned to include and report separately studies that investigated as biomarkers the ratio of proteases and their respective inhibitors (e.g. MMP‐2 (protease) and TIMP‐2 (inhibitor)).

The prognostic factor could be examined as a continuous or categorical variable and any cut‐off point was permitted.

We permitted any approach to obtaining samples from wounds, and any method of measurement of protease activity. We did not include measures of proteases in the blood or in tissue samples (biopsies).

We reported all unadjusted (simple) associations between the prognostic factor and healing and planned to report adjusted (multivariable) associations between the prognostic factor and healing, with details on any adjustment factors used, especially taking into consideration key adjustment factors of age and infection. No studies conducted multivariable analyses, and for some studies, we calculated unadjusted associations from raw data (e.g. mean and standard deviations of protease levels for healed and non‐healed wounds).

Types of outcomes

Selection of studies

Two review authors independently assessed the titles and abstracts of retrieved records against the inclusion criteria. We obtained all potentially relevant studies in full and two review authors independently assessed these for eligibility. We resolved any disagreements at each stage through discussion and, where appropriate, we consulted a third review author. We did not contact the study authors to resolve the uncertainty. Where studies were reported in multiple publications/reports, we obtained all publications. Whilst we included a study only once in the review, we extracted data from all reports to ensure we obtained all available relevant data. We collated multiple reports of the same study, so that each study rather than each report was the unit of interest in the review. The study selection process is illustrated in a PRISMA diagram, see Figure 1. All studies excluded after full‐text assessment are listed in a Characteristics of excluded studies table with their reasons for exclusion.

Figure 1

---

---

Study flow diagram.

Data extraction and management

We collected all association data reported between the prognostic factor and outcomes, with details on any adjustment factors used. Where necessary, we extracted raw data that would allow calculation of association measures.

We extracted and recorded data from included studies using an Excel‐based data extraction sheet, which we piloted initially on a few studies. We based the Items extracted on the CHARMS guidance (Moons 2014). One review author extracted data, which was then checked by a second review author. We did not contact the study authors if key data were missing from reports, because the overall risk of bias was too high.

We extracted the following data for the prespecified prognostic factors and outcomes in this review. We collected outcome data as described in the 'Types of outcome measures' section:

• country and setting in which study was conducted;

• study design;

• eligibility criteria;

• participant details;

• ulcer details, including duration of ulcer, ulcer size, number of ulcers and ulcer history, and ulcer severity (with the system used to classify this) as reported by the study authors;

• treatment details, including compression and debridement (including type and frequency);

• method of obtaining wound fluid for prognostic factor measurement, including duration of collection;

• method of measurement of protease activity (e.g. Gelatin zymography);

• details of each prognostic factor: type of protease/combination of factors, measurement of prognostic factor (including time of measurement), type of data (e.g. continuous/any cut‐off points and if so, whether they were predefined and what the justification was for that cut‐off point);

• details of each outcome: measurement, type of data (e.g. continuous/any cut‐off points);

• duration of follow‐up;

• type of analysis: explanatory/confirmatory (including the presence of a predefined protocol and study registration); logistic regression/Cox regression;

• any adjustment factors considered in the analysis;

• association statistics for each prognostic factor for primary and secondary outcomes (e.g. odds ratios (ORs), hazard ratios (HRs) and their CIs/variances/standard errors);

• loss to follow‐up and reasons.

We extracted minimal data where there were no association data reported.

Assessment of risk of bias in included studies

Two review authors independently appraised the included studies using a standardised approach. In the case of discrepancies, the review authors attempted to reach consensus; if necessary, a third review author resolved any disagreements. The review authors were not blinded to study authors, institution or journal of publication because this was not feasible.

We planned to assess risk of bias using an approach based on the Quality In Prognosis Studies (QUIPS) tool, which is appropriate for prognostic factor review questions (Hayden 2013; Hayden 2014), also drawing on ROBINS I tool (ROBINS‐I 2017), which assesses the risk of bias for non‐randomised intervention studies (Sterne 2016). However, because of the absence of multivariable analyses in all studies. we modified the approach (see below).

Measures of association

We planned to extract all unadjusted and adjusted measures of association from included studies and to convert effect sizes, as necessary, to avoid possible selection bias, thus allowing us to use data from as many studies as possible.

We had anticipated that results from multivariable analyses would have been reported as ORs, RRs and HRs and, if so, we would have used ORs as the common measure of the association, using RRs and HRs to estimate ORs at a particular time point (Symons 2002). However, none of the included studies reported adjusted measures for binary healing outcomes, but some studies provided sufficient raw data to allow us to calculate associations.

• Where both the healing outcomes and prognostic factors were reported as dichotomous variables (e.g. elevated and non‐elevated protease activities for complete healing), we formed 2 × 2 tables and analysed these data as the RR.

• Where the prognostic factors were reported as continuous variables and individual participant data (IPD) were given and the outcome was dichotomous, we conducted a univariate regression analysis in STATA (STATA 2013).

• Where there were summary data, we extracted the mean and standard deviation protease activity for each healing outcome state. To obtain an indication of possible associations between protease activity and healing, we conducted a t‐test to record associations between healing at the follow‐up time (as an independent variable) and protease levels at baseline (as the dependent variable) (i.e. we investigated the reverse association).

Where the outcomes were reported as continuous measures (e.g. change in ulcer size) and the prognostic factor was a dichotomous variable, we planned to analyse the regression coefficients with their standard errors or to conduct a simple t‐test. If both outcome and prognostic factor were continuous variables, we extracted (or calculated) Pearson correlation coefficients and P values and calculated CIs (Lowry 2018).

For consistency, we recalculated associations to be in the same direction, as necessary, with associations above 1 indicating better prognosis for positive binary outcomes (e.g. a healed wound). We did not contact study authors regarding missing or unusable data.

Unit of analysis issues

The prognostic factor (protease activity) and outcome (complete wound healing) were both considered at the ulcer level. A possible unit of analysis issue could have arisen when there was more than one ulcer per person and multivariable analysis was conducted with adjustment factors measured at the individual level (e.g. age). This represents clustered data and analyses should be conducted using hierarchical methods. Where studies included clustered data of this type, we planned to report this, noting whether data were analysed correctly, recording this as part of the risk of bias assessment. We planned to consider including such studies in any meta‐analysis, taking into account the associated risk of bias. No studies clearly included more than one wound per participant.

Dealing with missing data

We included studies that investigated the relationship between protease activity and healing regardless of whether there were missing data and even if limited evidence was provided about the size of the effect (e.g. if the factor was mentioned only as being ’non‐significant’ in the analyses). We did not contact study authors to attempt to retrieve any missing information.

Assessment of heterogeneity

We considered the clinical heterogeneity of included studies based on the population, measure of the prognostic factor, cut‐off points used, outcome measurement and methodological heterogeneity due to study design/potential biases. We planned to synthesise associations as appropriate within clinically relevant subgroups, grouping studies regardless of type of protease, duration of ulcer and presence of infection. We planned to examine the duration of ulcer and presence of infection in subgroup analyses where there was heterogeneity.

Assessment of reporting deficiencies

We planned to examine publication bias for each meta‐analysis, provided there were 10 or more studies, by visually examining asymmetry on funnel plots and testing for asymmetry at the 10% level, using Egger's test for HRs, and Peters' test for ORs (Debray 2018; Sterne 2011).

Data synthesis

Synthesising data in prognostic factor studies requires us to recognise the many different ways of reporting and analysing results, and, where possible, to request alternative data or to transform results to a common format (see 'Measures of association' section). We planned to consider the impact of data transformations using sensitivity analyses.

We had expected that most studies would present data in the format of dichotomous outcome data and continuous prognostic factor data (e.g. protease activity), which may have been dichotomised or categorised using cut‐off values .

After carrying out any appropriate transformations, where possible we grouped together studies with similar prognostic factor cut‐off points/similar analytical approaches, and represented the data on forest plots. We had planned to conduct meta‐analyses if valid data were available assessing associations between individual prognostic factors and an outcome of interest for sufficiently homogeneous subgroups of studies. We defined 'sufficiently homogeneous' subgroups according to population, measures of the prognostic factor and outcome measurement. However, we considered all data to be at too high a risk of bias to pool. We had planned to include in the forest plots details of any adjustment factors considered in the study analyses, but instead gave other relevant details in the footnotes.

We planned to combine the data for all proteases, regardless of the source or type of protease mainly because of a lack of evidence to inform stratification. For prognostic factors analysed as dichotomous measures, we planned additionally to take cut‐off points into consideration as described above. We planned to analyse separately ratios of proteases and their inhibitors.

If meta‐analysis had been conducted, we would have analysed HRs and ORs separately (at similar follow‐up points). We also planned to transform the measures, where possible, so that a single analysis (OR) could have been conducted. We planned to conduct meta‐regression analysis if there were more than 10 studies providing sufficient data (Berkey 1995). We planned to conduct meta‐analyses using STATA (StataCorp version 14) with a random‐effects restricted maximum likelihood (REML) meta‐analysis model, which accounts for any between‐study heterogeneity in the prognostic effect (Cornell 2014; Riley 2010). Such heterogeneity is common in prognostic factor studies. Unless the heterogeneity (as assessed above) was too extensive for appropriate pooling, we planned to summarise the meta‐analysis by the pooled estimate (the average prognostic factor effect), the Hartung‐Knapp 95% CI, the estimate of Tau² (between‐study variance) and a 95% prediction interval for the prognostic effect in a single population (Riley 2011).

'Summary of findings' table and GRADE assessment

We used an approach modified from the GRADE framework (Guyatt 2011a) to assess the certainty of the summarised evidence for each prognostic factor–outcome combination (Hayden 2014; Huguet 2013; Iorio 2015). We rated the overall strength of evidence as high, moderate, low or very low considering the phase of prognostic study (confirmatory/explanatory or exploratory), the within‐study risk of bias, the directness of evidence, heterogeneity, precision of effect estimates and risk of publication bias (Iorio 2015; Schünemann 2011a). We also considered two 'upgrading' factors, large effect and dose effect, although we noted that high risk of bias may artificially lead to large effects (see Appendix 4 for further details).

We did not present the main results of the review in formal 'Summary of findings' tables because the vast majority of the evidence was at high risk of bias and this, together with the study design, meant that the evidence was of very low certainty throughout. Instead we summarised the findings in the text. 'Summary of findings' tables present key information concerning the certainty of the evidence, the magnitude of the associations examined and the sum of the available data (Schünemann 2011b), and include an overall grading of the evidence. This defines the certainty of a body of evidence as the extent to which one can be confident that an estimate of effect or association is close to the true quantity of specific interest.

If it was not appropriate to combine results using a meta‐analysis (due to excess clinical heterogeneity), we planned to present the results qualitatively, considering the strength and consistency of results using the following schema:

• strong evidence of effect: consistent findings (defined as greater than 75% of studies showing the same direction of effect) in multiple low risk of bias studies;

• moderate evidence of effect: consistent findings in multiple high risk of bias or one study with low risk of bias (or both);

• limited evidence of effect: one study available;

• conflicting evidence of effect: inconsistent findings across studies;

• no effect: no association between participant expectations and the outcome of interest.

We planned to calculate absolute risk differences for the effect of the prognostic factor using estimates of baseline risk from the literature where possible.

Subgroup and sensitivity analyses

We planned to use sensitivity analyses to explore the impact of study level all‐domain risk of bias, first restricting the analysis to studies rated as having low risk of bias, and if this was not feasible, to restrict to low or moderate risk of bias.

If there was heterogeneity, we planned to investigate it using the following prespecified subgroup analyses, provided there were at least two studies per subgroup:

• baseline duration of ulcer (up to 24 weeks; 24 weeks or greater); duration may be a proxy for a non‐healing wound, in which protease activity may be associated differently with healing;

• presence or absence of infection.

We did not plan to conduct subgroup analyses by type of protease because this would have introduced too high a level of complexity for this review, but we recorded the type of protease measured.

We planned to consider subgroup or sensitivity analyses to explore the impact of types of measurement approaches for assessing prognostic factors.

This Methods section was based on the exemplar Cochrane prognosis review protocol for prognostic factors (Hayden 2014) and the general protocol template of the Cochrane Prognosis Methods Group (Cochrane Prognosis Methods Group 2018). In conducting the review, we carried out a number of modifications to the methods, as described above.