Why Science? | FDA Real-World Data, Real-World Evidence and Device Decisions

FDA describes real-world data as data relating to patient health status or delivery of health care that are routinely collected from a variety of sources. Real-world evidence is clinical evidence about the usage and potential benefits or risks of a medical product derived from analysis of RWD. Data are the ingredients; evidence is the result of a fit-for-purpose design and analysis.

The detective work in rwd and rwe are different objects is joyful when every clue keeps its address. Here, FDA describes real-world data as data relating to patient health status or delivery of health care that are routinely collected from a variety of sources. Real-world evidence is clinical evidence about the usage and potential benefits or risks of a medical product derived from analysis of RWD. Data are the ingredients; evidence is the result of a fit-for-purpose design and analysis. The guardrail is: A dataset does not become evidence merely because it is large or routinely collected. That is how routine records become an accountable scientific argument.

A strong science note separates observation from interpretation. Write separate definitions for the data source, the analytic method and the decision-relevant evidence output. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Next chapter

FDA issued the final device RWE guidance in December 2025 under docket FDA-2023-D-4395. It supersedes the final guidance issued on 31 August 2017. The current document clarifies how FDA evaluates RWD quality and provides updated recommendations for sponsors and reviewers considering RWE in device submissions.

For the december 2025 guidance is current and final, quantity is less interesting than alignment. FDA issued the final device RWE guidance in December 2025 under docket FDA-2023-D-4395. It supersedes the final guidance issued on 31 August 2017. The current document clarifies how FDA evaluates RWD quality and provides updated recommendations for sponsors and reviewers considering RWE in device submissions. The conclusion remains useful only while this limit is respected: Guidance describes current thinking; product-specific requirements and applicable law still govern.

A strong science note separates observation from interpretation. Build a version receipt with title, final status, month, docket and the superseded guidance date. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

The question should specify the device, intended use, population, exposure or treatment, comparator where relevant, outcome, time window and regulatory purpose. This framework reveals whether a source can supply the needed variables and follow-up. Starting with a convenient database can tempt analysts to reshape the question around whatever fields happen to exist.

Picture a regulatory question comes before a data search as a chain of custody for meaning. Data move through people, systems and code before becoming a result. In this chapter, The question should specify the device, intended use, population, exposure or treatment, comparator where relevant, outcome, time window and regulatory purpose. This framework reveals whether a source can supply the needed variables and follow-up. Starting with a convenient database can tempt analysts to reshape the question around whatever fields happen to exist. Keep the following boundary attached: One RWD source may be fit for one question and unfit for another.

A strong science note separates observation from interpretation. Write a structured question and mark every element that must be observed or derived in the data. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

Potential sources include electronic health records, medical claims, product or disease registries, patient-generated data and data gathered from sources such as digital health technologies. Each source reflects a workflow and purpose. Claims may capture billed services; a registry may capture specialised clinical detail; device data may capture frequent measurements.

The detective work in rwd sources have different strengths is joyful when every clue keeps its address. Here, Potential sources include electronic health records, medical claims, product or disease registries, patient-generated data and data gathered from sources such as digital health technologies. Each source reflects a workflow and purpose. Claims may capture billed services; a registry may capture specialised clinical detail; device data may capture frequent measurements. The guardrail is: No source is complete simply because it is digital. That is how routine records become an accountable scientific argument.

A strong science note separates observation from interpretation. For two sources, list the workflow that creates the data and the important clinical event that workflow might miss. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

Relevant RWD contain the population, exposure, comparator, outcomes, covariates and follow-up needed for the proposed analysis. They should be sufficiently representative of the intended use and provide enough observations for meaningful inference. Relevance is therefore relational: it is judged against the regulatory question, not as a permanent property of a database.

For relevance asks whether the data can answer the question, quantity is less interesting than alignment. Relevant RWD contain the population, exposure, comparator, outcomes, covariates and follow-up needed for the proposed analysis. They should be sufficiently representative of the intended use and provide enough observations for meaningful inference. Relevance is therefore relational: it is judged against the regulatory question, not as a permanent property of a database. The conclusion remains useful only while this limit is respected: Representativeness and sample size do not cure missing key variables.

A strong science note separates observation from interpretation. Create a variable-to-source map and label each item direct, derived, linked, incomplete or unavailable. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

Reliability concerns how data are collected, curated, transformed and maintained. Important evidence may include data-entry procedures, definitions, validation, quality checks, completeness, consistency and auditability. A familiar database name is not enough; reviewers need to understand the process that produced the analytical dataset.

Picture reliability asks whether the data can be trusted as a chain of custody for meaning. Data move through people, systems and code before becoming a result. In this chapter, Reliability concerns how data are collected, curated, transformed and maintained. Important evidence may include data-entry procedures, definitions, validation, quality checks, completeness, consistency and auditability. A familiar database name is not enough; reviewers need to understand the process that produced the analytical dataset. Keep the following boundary attached: A validated source can still be misused through incorrect extraction or transformation.

A strong science note separates observation from interpretation. Trace one critical outcome from clinical event to source record, extract, transformation and analysis variable. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

A fit-for-purpose assessment explains why the selected data and methods are adequate for the particular regulatory use. It brings relevance, reliability, study design, analysis and uncertainty into one argument. The threshold should reflect the decision’s importance and the consequences of error.

The detective work in fitness for purpose joins relevance and reliability is joyful when every clue keeps its address. Here, A fit-for-purpose assessment explains why the selected data and methods are adequate for the particular regulatory use. It brings relevance, reliability, study design, analysis and uncertainty into one argument. The threshold should reflect the decision’s importance and the consequences of error. The guardrail is: Fitness is specific to the purpose and cannot be transferred automatically to a new claim. That is how routine records become an accountable scientific argument.

A strong science note separates observation from interpretation. Write a one-paragraph fitness argument and a separate paragraph describing the strongest reason it could fail. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

Data provenance records origin, ownership, collection context, time, versions and transformations. When several sources are linked, provenance should show matching rules and what happens when records conflict. Without it, an analyst may reproduce code yet still be unable to reconstruct the data that entered the model.

For provenance keeps the evidence route visible, quantity is less interesting than alignment. Data provenance records origin, ownership, collection context, time, versions and transformations. When several sources are linked, provenance should show matching rules and what happens when records conflict. Without it, an analyst may reproduce code yet still be unable to reconstruct the data that entered the model. The conclusion remains useful only while this limit is respected: A reproducible script cannot compensate for an unreconstructable source dataset.

A strong science note separates observation from interpretation. Build a lineage diagram from source systems through linkage and curation to the final analysis table. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

Linking registry, claims, electronic record or device data may create a richer patient journey. It can also introduce missed matches, false matches and unequal linkage across groups. The protocol should describe identifiers, deterministic or probabilistic rules, evaluation of linkage quality and handling of unmatched records.

Picture linkage can add insight and error as a chain of custody for meaning. Data move through people, systems and code before becoming a result. In this chapter, Linking registry, claims, electronic record or device data may create a richer patient journey. It can also introduce missed matches, false matches and unequal linkage across groups. The protocol should describe identifiers, deterministic or probabilistic rules, evaluation of linkage quality and handling of unmatched records. Keep the following boundary attached: Linked records are not automatically more accurate than their component sources.

A strong science note separates observation from interpretation. Report match rate, possible false matches, unmatched characteristics and a sensitivity analysis for linkage uncertainty. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

An outcome may be directly recorded, coded, adjudicated or derived from several fields. The definition should specify code sets, windows, repeated events and validation evidence. A hospitalisation code, laboratory result and patient-reported symptom measure different things, even when all are used as proxies for a broad concept such as deterioration.

The detective work in outcome definitions need clinical and data logic is joyful when every clue keeps its address. Here, An outcome may be directly recorded, coded, adjudicated or derived from several fields. The definition should specify code sets, windows, repeated events and validation evidence. A hospitalisation code, laboratory result and patient-reported symptom measure different things, even when all are used as proxies for a broad concept such as deterioration. The guardrail is: A convenient proxy should not be described as the clinical concept without validation. That is how routine records become an accountable scientific argument.

A strong science note separates observation from interpretation. Write the clinical concept, operational definition and evidence that the definition detects the intended event. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

RWD often contain missing values because care was delivered elsewhere, a test was not ordered, a field was optional or a device stopped transmitting. Those reasons may relate to health, access or outcome. Analysts should describe missingness, investigate patterns, justify handling methods and assess how conclusions change under plausible alternatives.

For missingness is a pattern to investigate, quantity is less interesting than alignment. RWD often contain missing values because care was delivered elsewhere, a test was not ordered, a field was optional or a device stopped transmitting. Those reasons may relate to health, access or outcome. Analysts should describe missingness, investigate patterns, justify handling methods and assess how conclusions change under plausible alternatives. The conclusion remains useful only while this limit is respected: Complete-case analysis can be biased when missingness relates to exposure or outcome.

A strong science note separates observation from interpretation. Create a missingness table by study group and time, then list clinical reasons for the largest gaps. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

Device use in routine care is influenced by clinician choice, disease severity, access, preference and other factors that may also affect outcomes. Design and analysis can address measured confounding through restriction, matching, weighting, stratification or modelling, but unmeasured confounding can remain. A clear causal diagram can expose assumptions before estimation.

Picture confounding threatens non-random comparisons as a chain of custody for meaning. Data move through people, systems and code before becoming a result. In this chapter, Device use in routine care is influenced by clinician choice, disease severity, access, preference and other factors that may also affect outcomes. Design and analysis can address measured confounding through restriction, matching, weighting, stratification or modelling, but unmeasured confounding can remain. A clear causal diagram can expose assumptions before estimation. Keep the following boundary attached: Adjustment cannot remove a confounder that was not measured well enough.

A strong science note separates observation from interpretation. Draw a causal diagram and distinguish confounders, mediators, colliders and variables unavailable in the source. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

A transparent plan defines eligibility, exposure, comparator, outcomes, follow-up, estimand, covariates, missing-data methods and primary analyses before results are known. Registration, version control and dated amendments help separate planned work from exploration. Deviations should be explained rather than quietly absorbed into the final report.

The detective work in protocol and analysis should be specified early is joyful when every clue keeps its address. Here, A transparent plan defines eligibility, exposure, comparator, outcomes, follow-up, estimand, covariates, missing-data methods and primary analyses before results are known. Registration, version control and dated amendments help separate planned work from exploration. Deviations should be explained rather than quietly absorbed into the final report. The guardrail is: Prespecification reduces selective analysis but does not guarantee a sound design. That is how routine records become an accountable scientific argument.

A strong science note separates observation from interpretation. Freeze a protocol hash or version, then log every change with date, reason and whether outcomes were known. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

RWE conclusions may depend on outcome definitions, exposure windows, confounding control, missingness, linkage or model form. Sensitivity analyses should target these assumptions rather than produce decorative extra models. Negative controls or quantitative bias analyses may sometimes help reveal residual error when scientifically justified.

For sensitivity analyses test fragile assumptions, quantity is less interesting than alignment. RWE conclusions may depend on outcome definitions, exposure windows, confounding control, missingness, linkage or model form. Sensitivity analyses should target these assumptions rather than produce decorative extra models. Negative controls or quantitative bias analyses may sometimes help reveal residual error when scientifically justified. The conclusion remains useful only while this limit is respected: Agreement across similar models does not test assumptions they all share.

A strong science note separates observation from interpretation. Name the assumption most likely to reverse the conclusion and design an analysis that actually perturbs it. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

FDA describes RWE as potentially useful across the device life cycle and in regulatory submissions, including premarket and postmarket contexts. Examples may involve supporting safety or effectiveness questions, expanding information about use or monitoring performance. The evidentiary role should be explicit, because supportive evidence and primary evidence carry different burdens.

Picture rwe can support different device decisions as a chain of custody for meaning. Data move through people, systems and code before becoming a result. In this chapter, FDA describes RWE as potentially useful across the device life cycle and in regulatory submissions, including premarket and postmarket contexts. Examples may involve supporting safety or effectiveness questions, expanding information about use or monitoring performance. The evidentiary role should be explicit, because supportive evidence and primary evidence carry different burdens. Keep the following boundary attached: Past acceptance of RWE for one decision does not guarantee acceptance for another.

A strong science note separates observation from interpretation. State whether the RWE is primary, supportive or contextual, and identify the exact decision it is intended to support. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

A strong submission documents sources, governance, protocol, variable definitions, linkage, curation, code or algorithms, quality checks, analyses, deviations and limitations. Tables should keep denominators, dates and populations visible. Reproducibility means another qualified analyst can follow the route, not merely rerun an unexplained file.

The detective work in transparency makes review and reproduction possible is joyful when every clue keeps its address. Here, A strong submission documents sources, governance, protocol, variable definitions, linkage, curation, code or algorithms, quality checks, analyses, deviations and limitations. Tables should keep denominators, dates and populations visible. Reproducibility means another qualified analyst can follow the route, not merely rerun an unexplained file. The guardrail is: Confidentiality constraints may limit sharing, but they do not justify opaque methods. That is how routine records become an accountable scientific argument.

A strong science note separates observation from interpretation. Prepare a reviewer map linking every main claim to dataset version, code version, output and limitation. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

The receipt should record regulatory question, data owners, source versions and dates, population, variables, provenance, linkage, quality assessment, protocol, analysis code, missingness, bias controls, sensitivity results, findings, limits and FDA source. It becomes the memory of what the evidence can support when people and systems change.

For a complete rwe receipt protects future reuse, quantity is less interesting than alignment. The receipt should record regulatory question, data owners, source versions and dates, population, variables, provenance, linkage, quality assessment, protocol, analysis code, missingness, bias controls, sensitivity results, findings, limits and FDA source. It becomes the memory of what the evidence can support when people and systems change. The conclusion remains useful only while this limit is respected: A receipt preserves the analysis boundary; it does not make the source fit for a new purpose.

A strong science note separates observation from interpretation. Hand the receipt to a new analyst and ask them to identify what must be revalidated before answering a related question. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter