Building an external control arm (ECA) for a wound care trial is a data engineering problem disguised as a statistical one. The methodology, propensity matching, IPTW, SMD reporting, is well documented in oncology and cardiology RWE literature. What’s missing is guidance on the wound-specific failure points: fragmented care settings, inconsistent wound area measurement, and CTP/CAMP product switching that doesn’t show up in a typical claims dataset. This guide assumes you already have a locked trial protocol with defined inclusion criteria and outcome windows, and access to a real-world dataset with wound-level granularity, etiology, size, depth, treatment setting, not just diagnosis codes. If either of those isn’t in place yet, the steps below won’t hold up under FDA or MAC scrutiny.
Step 1: Define the Regulatory Objective and Target Estimand
Before pulling a single record, determine exactly what regulatory or evidentiary job the ECA is doing. An EU MDR post-market clinical follow-up (PMCF) obligation, an FDA label expansion submission, and a payer coverage dossier each carry different evidentiary thresholds. FDA’s expectations for real-world evidence supporting a labeling claim are considerably more rigorous than what a Medicare Administrative Contractor (MAC) will accept in an LCD reconsideration package. Confusing the two leads teams to either over-engineer a payer submission or under-engineer an FDA one.
Once the objective is set, lock the estimand. This means specifying the exact outcome (for example, time to 100% wound closure at 12 weeks), the population (DFU versus VLU, wound duration and size bounds), and the comparison structure, all before touching any data source. Post-hoc outcome switching, where a team pivots from closure rate to wound area reduction after seeing which metric produces a cleaner comparison, is one of the fastest ways to have an ECA dismissed by a reviewer. Write the estimand into the statistical analysis plan (SAP) and treat it as immutable once data extraction begins.
Finally, decide the ECA’s structural role in the evidence package. Is it the primary comparator for a single-arm trial, effectively standing in for a randomized control group, or is it a sensitivity analysis running alongside a randomized comparator to test robustness? A single-arm trial leaning entirely on an ECA needs a substantially more defensible matching strategy and broader sensitivity testing than an ECA used as supporting context next to an RCT. Clarifying this upfront determines how much statistical firepower Steps 4 through 6 actually need, and it shapes how reviewers at FDA or in a comparative effectiveness study will weigh the result. Teams that skip this step tend to discover, three months into analysis, that they built the wrong estimand for the objective they were actually pursuing.
Step 2: Select a Fit-for-Purpose Real-World Data Source
Not all real-world wound care data is usable for an ECA, regardless of patient volume. The first filter is wound-level granularity: does the dataset capture etiology, wound area and depth at baseline, prior treatment failures, and product-level CTP/CAMP application history, or does it only offer ICD-10 codes and CPT-coded procedures? Claims-only sources are generally insufficient on their own because they can’t reconstruct wound size trajectories or confirm the specific product applied at each visit. A registry or EHR-derived dataset with structured wound measurements, ideally supported by AI-based wound surface area measurement, is far more defensible as an ECA source.
Second, confirm care-setting coverage matches your trial population. Utilization patterns for CTPs and CAMPs differ meaningfully across hospital-based wound centers, skilled nursing facilities, and home health visits, both in product selection and in documented adherence to offloading or compression protocols. If your trial enrolled patients exclusively from hospital-based outpatient wound centers, an ECA drawn primarily from SNF data will introduce a setting-driven confound before you’ve matched a single covariate. Intellicure Analytics’ Wound Care Industry Dashboard, for instance, tracks product utilization across hospital-based wound centers, SNFs, and home visits separately, which lets teams select or stratify by the setting that actually mirrors the trial population rather than blending settings and hoping the imbalance washes out in matching.
Third, check data recency and refresh cadence. A dataset with a documented “as of” date and a predictable update schedule is easier to defend under review than a static, one-time extract of unknown vintage. Reviewers, whether at FDA or a MAC, will ask when the data was pulled and whether practice patterns have shifted since. Build that documentation into your data provenance narrative from the outset rather than reconstructing it later.
Step 3: Mirror the Trial’s Eligibility Criteria in the Real-World Cohort
The real-world cohort has to be built to the trial’s inclusion and exclusion criteria, not adapted to whatever the data happens to support. Start by translating every criterion, wound duration, size range, ABI or non-invasive arterial assessment thresholds, exclusions like uncontrolled diabetes or active osteomyelitis, into a discrete, extractable data field. Some criteria won’t map cleanly. If the trial excluded patients with HbA1c above a certain threshold and your dataset doesn’t capture lab values reliably, flag that as an explicit limitation rather than approximating it with a proxy and presenting it as equivalent.
Index date logic matters as much as the criteria themselves. Use the same anchor the trial used, typically first CTP or CAMP application, to define time zero in the real-world cohort. Getting this wrong introduces immortal time bias: patients who had to survive with an open wound long enough to receive treatment get credited with follow-up time from an earlier point than they should, artificially inflating the apparent effectiveness of the treated arm relative to the comparator.
The most common mistake at this stage is loosening eligibility criteria to pad sample size, then planning to “adjust for it later” with covariates. This is backwards. Broadening the cohort first and matching second increases confounding risk rather than reducing it, because you’re now asking a statistical model to correct for differences that better cohort selection would have avoided entirely. If sample size is genuinely insufficient after applying strict criteria, that’s a finding to report, not a problem to paper over with a wider net. It’s also a signal to reconsider the data source itself; a registry-based wound care research platform with broader multi-site coverage may resolve the volume problem without loosening criteria.
Step 4: Extract and Harmonize Baseline Covariates
Build a covariate table before any matching work begins. At minimum, this should cover demographics, wound characteristics (etiology, size, depth, duration, anatomic location), relevant comorbidities (peripheral arterial disease, chronic kidney disease, HbA1c where captured), and prior treatment history including failed conservative care and prior CTP/CAMP exposure. This table becomes the backbone of both the matching model and the balance reporting in Step 5.
If you’re blending sources, say, a clinical registry with EHR-derived extracts, unit and coding harmonization is not optional. Wound area measurement is the classic failure point here: one source may record length times width as a proxy for area, another may use planimetry or AI-based wound surface area measurement, and a third may round to the nearest centimeter. These aren’t interchangeable, and treating them as such introduces measurement noise that can swamp a real treatment effect, particularly for outcomes defined around percentage area reduction. Standardize to a single measurement methodology before merging, or stratify analyses by measurement source and test whether results hold across both.
Missingness needs to be assessed field by field, not just at the dataset level. As a rule of thumb, a covariate missing in fewer than 5% of records can often be handled with simple imputation and a sensitivity check. A covariate missing in more than 20% of records, HbA1c is a frequent offender in real-world wound datasets, needs an explicit imputation strategy documented in the SAP: multiple imputation, missing-indicator methods, or exclusion with justification. Reviewers will ask how missing data was handled; “we didn’t have it” is not an acceptable answer if the field was material to the matching model. Document the missingness pattern itself, since data missing systematically by care setting or by site is a different problem than data missing at random.
Step 5: Apply a Matching or Weighting Method
The choice between propensity score matching and inverse probability of treatment weighting depends largely on sample size and covariate overlap between the trial and real-world populations. Propensity score matching, 1:1 or 1:n, is more intuitive to explain to non-statistical reviewers and produces a cleaner, smaller matched cohort, but it discards unmatched patients, which can be costly if the real-world pool is already limited after Step 3’s strict eligibility filtering. IPTW retains the full sample and can be more statistically efficient, but it’s more sensitive to extreme weights when overlap is poor and harder to communicate to a lay reviewer or payer committee.
Before finalizing either approach, check common support, the degree of overlap between the propensity score distributions of the trial arm and the real-world candidate pool. If the two populations occupy substantially different regions of the covariate space, for example, the trial enrolled younger patients with smaller wounds than what’s available in the real-world dataset, no amount of statistical adjustment fixes that. Trimming to the region of common support, or acknowledging that the ECA can’t answer the question for the excluded subgroup, is more defensible than forcing a match across non-overlapping populations.
Once matching or weighting is applied, report standardized mean differences for every covariate in the table, both pre- and post-adjustment. The conventional threshold is an SMD below 0.1 post-matching to consider a covariate adequately balanced; values above that signal residual imbalance that needs to be addressed, either through refined matching parameters or explicit acknowledgment as a limitation. Present this as a table, not a summary sentence. Reviewers evaluating an ECA, whether for an FDA submission or a comparative effectiveness study, will look at the covariate balance table before they look at the topline effect estimate, because balance is what determines whether that effect estimate means anything.
Step 6: Run Sensitivity Analyses and Bias Assessments
A single matching specification is never sufficient to support a submission-grade ECA. Rerun the analysis with alternative matching ratios (1:1 versus 1:2 or 1:3), different caliper widths, and alternative covariate sets that add or drop borderline variables. If the treatment effect estimate shifts materially across these variations, that instability is itself a finding, and it usually means the original result was more an artifact of modeling choices than a robust signal. Stability across reasonable specifications is what gives a reviewer confidence the result isn’t a product of researcher degrees of freedom.
Where feasible, run a negative control outcome analysis, testing the same matched cohorts against an outcome the treatment shouldn’t plausibly affect. If the ECA shows a “significant” difference on an outcome with no biological reason to differ between arms, that’s evidence of unmeasured confounding contaminating the comparison, and it should prompt a hard look at what’s missing from the covariate table before the primary result is trusted.
Document limitations explicitly rather than burying them. No ECA is equivalent to a randomized trial, and unmeasured confounders specific to wound care, patient adherence to offloading in DFU cases, compression adherence in VLU cases, caregiver support in home health settings, are rarely captured in any real-world dataset. State this plainly in the limitations section rather than implying equivalence to randomization. FDA reviewers and journal referees are more likely to accept an ECA that owns its limitations than one that oversells its rigor.
Step 7: Package the Analysis for Submission or Publication
The final package needs to stand on its own for a reviewer who wasn’t involved in the analysis. Structure it around data provenance (source, refresh cadence, care-setting coverage), a cohort derivation flow diagram showing exactly how the analytic population was reduced from the raw dataset through eligibility filtering and matching, full covariate balance tables (pre- and post-match SMDs), and the complete set of sensitivity analysis results from Step 6. This structure aligns with FDA’s final guidance, Use of Real-World Evidence to Support Regulatory Decision-Making for Medical Devices, issued December 18, 2025, which supersedes the 2017 version; citing the outdated 2017 document undermines a submission’s credibility.
If the ECA will also support a peer-reviewed publication alongside a PMCF filing, check the target journal’s reporting requirements early, many now expect adherence to the RECORD statement, STROBE, or similar reporting guidelines for real-world evidence studies, and build the analysis documentation to satisfy both the regulatory and editorial audience simultaneously rather than writing two separate versions after the fact.
Retain the raw extraction logic, cohort-building code, and matching scripts in a version-controlled, auditable form. Reviewers are increasingly requesting reproducibility documentation for RWE-based comparisons, and being able to hand over exact code rather than a narrative description of methods shortens review cycles considerably. This is also where a documented, membership-based data platform has an advantage over a one-off data pull: the extraction logic and dataset version are traceable after the fact, which matters if a reviewer comes back eighteen months later with questions.
Confirming the ECA Holds Up Before Submission
An ECA that looks statistically clean but hasn’t been pressure-tested against current FDA real-world evidence guidance is a liability, not an asset, once it reaches review. Before submission, walk back through every covariate imbalance flagged in Step 5, every sensitivity result from Step 6, and confirm each is documented rather than smoothed over in the narrative. The teams that get pushback aren’t usually the ones with imperfect data; they’re the ones who didn’t disclose the imperfections. Intellicure Analytics supports manufacturers and researchers building these comparisons with wound-level real-world data spanning hospital-based wound centers, SNFs, and home health, along with comparative effectiveness and PMCF services designed for this exact evidentiary purpose. Learn more about our services