Ask most development teams when real-world evidence enters a programme and you get a reassuringly tidy answer: it's the package you assemble for submission, and the registry you commit to once the drug is approved and the agency wants long-term safety. RWE arrives late: useful and increasingly expected, but bolted on near the finish line.
That picture is wrong in a specific and expensive way. RWE has a usable entry point at every stage of a drug development programme, from sizing a disease before a target is chosen to expanding a label years after launch, and the entry points that change the most sit before the protocol is ever locked.
Look at the statutory seed. When Congress first wrote real-world evidence into US law through the 21st Century Cures Act §3022 (codified at 21 U.S.C. §355g, enacted 13 December 2016), it scoped the term narrowly: to "help to support the approval of a new indication for a drug" and to "help to support or satisfy postapproval study requirements" [1]. Two uses, both of them late: the law planted its seed at the far end of the pipeline, and the map that grew from it is far wider.
Most teams still behave as though the seed were the whole map. RWE gets pulled off the shelf at submission, or worse, when a trial is already on fire and someone needs a comparator by Friday. Call it the "break-glass reflex": evidence treated as an emergency instrument rather than a design input. It works, sometimes, but it leaves the richest ground on the map unwalked, because nobody goes there until the walls are already smoking. (For the trial-on-fire case, see how RWE rescues a programme gone sideways.)
Here's the asymmetry that runs through the whole thing. The named regulatory wins cluster late: external control arms, label expansions, safety signals, HTA dossiers. Those nodes are the charted coastline, mapped in detail because that's where everyone has already sailed. The upstream nodes, the ones that shape what you actually build, are the faintly drawn interior, unvisited rather than empty.
You don't have to take my word for where the nodes sit, because a regulator drew them. The EMA's Reflection paper on the use of real-world data in non-interventional studies (EMA/99865/2025, 3 April 2025) walks the uses in order, from "describing disease epidemiology (incidence, prevalence, risk factors and progression)" through "Supporting the planning and feasibility assessment of a clinical study by characterising patients, exposure and endpoints" to investigating post-marketing utilisation, safety and effectiveness [2]. A regulator toured the entire lifecycle. Most sponsors visit two stops on it. So let's walk inland.
Let's be honest about the limit first. RWE does not discover drug targets. There's no named case of a real-world dataset producing a de novo target that went on to approval, and anyone who tells you otherwise has something to sell. What RWE does upstream is quieter: it informs target rationale, sizes the addressable disease, and surfaces repurposing signals from how medicines already get used.
The concrete use is disease sizing. Before you sign off a target rationale, you can characterise the disease from claims and EHR rather than inheriting a KOL's back-of-envelope figure. Neuroendocrine tumour prevalence, for one, has been characterised from large claims databases such as MarketScan and PharMetrics, per Dagenais and colleagues (2022) citing the primary source [3]. It sounds dull. It's also the difference between a programme sized on data and one sized on a number someone half-remembers from a conference.
Monday test: before the target rationale is signed off, size the addressable population and map the current treatment pathway from real-world data. More on how biotechs reduce development risk with RWE early.
The single most useful thing you can know before you design anything is your disease's untreated trajectory. Natural history is the counterfactual you'll eventually be judged against, and it exists whether or not you've looked at it.
Take amyotrophic lateral sclerosis. A natural-history cohort of 175 patients carrying the A4V-SOD1 mutation showed a median survival of roughly 1.2 years, per Dagenais 2022 (again, a secondary figure the review draws from the primary) [3]. That's the yardstick that later tells you whether an effect you observe is your drug or the disease running its normal course. On the endpoint side, the same review points to haemoglobin as a validated surrogate in sickle cell disease, established from real-world data long before any given trial locks it in [3].
Many years ago I worked on an early-phase rare-disease programme where a cluster of adverse events surfaced mid-trial, and the whole question, drug or disease, hung on whether we held a natural-history baseline solid enough to compare against; we did, barely, and it settled the argument in an afternoon rather than a submission cycle. Establish the counterfactual before the readout, not after. This is internal-decision territory, not landmark-approval territory. And that's the point: the value is real, and almost entirely un-banked.
This is where upstream RWE earns its keep hardest, and where the fewest teams look. It comes in two flavours.
The first is feasibility and sample-size modelling against real patients. In an analysis Dagenais 2022 cites, the Trial Pathfinder tool run across 61,000 NSCLC patients showed that loosening over-strict eligibility criteria could cut the required sample size "by at least 40%", with "time savings of at least 6 months" (secondary figures, attributed to the primary the review draws on) [3]. Forty per cent fewer patients, and six months of runway. For a lean biotech, that difference decides whether you reach the next raise or stall short of it. The EMA names the use directly: RWD supports "the planning and feasibility assessment of a clinical study by characterising patients, exposure and endpoints" [2]. A registry like the CINRG Duchenne natural-history study (NCT00468832, n=551) exists to do exactly that, its registered aims including validation of the 6MWT and NSAA as endpoints and sample-size calculation for future trials [17].
The second flavour is RWD woven into the randomised design itself, a different animal from an external control. Registry-based and pragmatic RCTs route the randomisation through routine-care infrastructure: TASTE (NCT01093404, n=7,243) nested a STEMI trial inside the Swedish SCAAR and SWEDEHEART registries; ADAPTABLE (NCT02697916, n=15,076) ran aspirin dosing on EHR and PCORnet; the Salford Lung Study (NCT01551758, n=2,802) tested a COPD therapy in routine primary care with linked records [17].
Monday test: run the feasibility question against real patient data before the protocol locks, and decide whether a pragmatic or registry-based design actually fits your question. On not over-buying data while you do it, see the minimum viable evidence framework.
Free download
The RWE Briefing Document Template
The section-by-section structure for the RWE part of a regulatory briefing, built around the questions reviewers actually ask.
Get the template →Choose an endpoint the real world can't reliably capture and you'll pay for it later, when that endpoint has to survive as an external-control or label-expansion measure. Far better to find out early.
Griffith and colleagues (2019), working in the Flatiron oncology dataset, put numbers on the trap. Trying to lift a trial-style endpoint straight into real-world records, they found "a RECIST-based approach was not feasible: cancer progression could be ascertained for 23% (6/26 patients)" [4]. A clinician-anchored real-world PFS of about 5.5 months was measurable, and it tracked overall survival at a Spearman's rho of 0.65–0.66 [4]. Real-world endpoints are usable once validated and dangerous when assumed. Rubbish in, rubbish out absolutely applies.
Monday test: before you propose an endpoint for any real-world-facing purpose, confirm it's ascertainable in the data you'll actually have, not the data you wish you had.
Back on the charted coastline now, at the node teams do reach for. It's real, but it's conditional, and pretending otherwise gets programmes into trouble.
The worked example: in Arondekar and colleagues' 2022 review, avapritinib's real-world "ORR of 0%–5%" was "contextualized with the ORR of 84% from the phase I trial" and thereby "helped demonstrate the clinical benefit of avapritinib" [5]. A comparator that poor made a trial result that strong legible to the agency. But note the scale: across 133 original oncology approvals, only 11 used RWE, and "none predated 2017" [5].
Late, and rare.
Rare disease leans on this node harder. Vaghela and colleagues (2024) found that of 20 rare-disease applications using RWD, 14 (70%) drew on natural history including "registry-based/retrospective historical controls", and nine of the twenty (45%) received positive FDA feedback, mainly where "a significantly large effect size" overcame the inherent bias [6]. The named cases are instructive: vutrisiran's HELIOS-A used the APOLLO placebo arm as an external control; Zolgensma leaned on SMA natural history; Voxzogo used a matched achondroplasia cohort, 559 narrowed to 360 [6]. The trial records make it tangible: cerliponase alfa in CLN2 (NCT01907087, n=24) measured efficacy against a natural-history control, and Zolgensma's STR1VE (NCT03306277, n=22) ran single-arm against untreated SMA1 natural history [17].
That said, the conditional half isn't optional reading. Alipour-Haris and colleagues (2024) found 13 use cases where RWE "was not considered supportive/definitive" on account of design issues [7]. The node is genuinely open; acceptance is earned, not granted.
Monday test: if you're running or planning a single-arm trial, build the external comparator now, to the standard the agency will actually apply, not the one you wish it would. Our external-control-arm checklist walks the FDA's expectations line by line.
By submission, the late nodes converge into one job: present RWE to the evidentiary bar, in the format the agency expects. Alipour-Haris (2024) catalogued 85 regulatory RWE use cases: 59 (69.4%) supporting an original marketing application, 24 (28.2%) a label expansion, and 2 (2.4%) a label modification, with 42 deployed to support single-arm trials [7]. The named exemplars run the range: Zalmoxis (a registry external control), Erbitux (an EHR-based label modification), Intelence (a registry label expansion) [7].
What you must know cold is which guidance governs which use, and its status, because a wrong assumption there is expensive. Final: "Considerations for the Use of Real-World Data and Real-World Evidence" (August 2023), "Submitting Documents Using Real-World Data and Real-World Evidence" (September 2022), "Assessing Electronic Health Records and Medical Claims Data" (July 2024), and "Assessing Registries" (December 2023) [12]. Still in draft, and this is where teams get caught: "Externally Controlled Trials" (February 2023) and "Non-Interventional Studies" (March 2024) [12]. Above all of it sits the Framework for FDA's Real-World Evidence Program (December 2018) [11]. If your pivotal contribution leans on an external control or a non-interventional study, you're standing on draft guidance.
Monday test: map each RWE element in your package to the specific final guidance that governs it, and flag the ones sitting under draft guidance for early agency discussion. On using RWE to support regulatory strategy generally; and on what "regulatory-grade" RWE actually demands, which is not the brand name of your database.
Three sub-nodes, and the highest-volume, most-accepted use of RWE anywhere lives in the first.
Safety. Pharmacovigilance is where RWE is least controversial and most used. Crisafulli and colleagues (2025) give the worked example: a DOAC study of "more than 1 million new users" found rivaroxaban-associated severe uterine bleeding at "hazard ratios that ranged from 1.19 to 1.34", which "contributed to a class-wide labelling change for DOACs" [8]. The infrastructure is industrial: FAERS alone holds "over 29.6 million drug-focused ICSRs" across "20.1 million unique cases", and systems like FDA Sentinel, DARWIN EU and VigiBase run continuously [8]. Mind the boundary of the newest guidance: ICH M14 (draft, Step 2b, 21 May 2024) covers RWD for safety pharmacoepidemiology only, not effectiveness, trial design or label expansion [13]. Real commitments already sit on the register, among them the ARISER post-authorisation safety study for Zolgensma (NCT06019637, real-world safety followed to 15 years) and a tisagenlecleucel real-world CAR-T registry in Korea (NCT06785818) [17].
Market access. Payers were early, pragmatic adopters. Boss and colleagues (2026) reviewed 379 CDA-AMC reimbursement submissions from 2017 to 2022 and found RWE in roughly a third (n=145), rising to about half in oncology, often used to "fill gaps in economic model inputs (27 percent)" [9]. NICE's RWE framework (ECD9, 23 June 2022) names the uses plainly: to "design, populate and validate economic models" and to "assess the applicability of clinical trials to patients in the NHS" [14].
Label expansion. This is the late win teams do associate with RWE, and it pays to be honest about its size. Deng and colleagues (2025) found that among 218 labelling expansions, "25.2%… included or likely included RWE", most commonly in oncology (43.6%) [10]. Two landmark effectiveness-via-RWD approvals are worth naming: Ibrance (palbociclib) in male breast cancer (4 April 2019), "supported by real-world data from electronic health records and insurance claims" per the FDA approval summary [15]; and Prograf (tacrolimus) for lung-transplant effectiveness (16 July 2021), granted on observational data benchmarked against SRTR natural history [16]. Neither ran a fresh interventional trial to get there.
Monday test: the post-approval nodes are commitments you can pre-empt. Design the safety registry and the economic-model inputs while the pivotal trial is still enrolling, not after the agency asks.
Let me put the strongest version of the counterargument, because this audience will raise it anyway. Isn't this just a consultant's everything-everywhere map, an "RWE can help here" sticker slapped over every box in the lifecycle? And doesn't the payer settle the matter? In Boss's interviews the boundary is blunt: RWE can "complement… but never supplant" RCT data [9]. The hard numbers seem to agree. Deng found RWE in only a quarter of 218 label expansions [10]; Arondekar found it in just 11 of 133 original oncology approvals [5]. On that reading, RWE is a marginal add-on, and a map that makes it look ubiquitous is overselling.
Fair. But the objection conflates two things the numbers blur. RWE for any purpose, safety, feasibility, disease sizing, is common and rising. RWE as pivotal efficacy evidence is rare and conditional, and always will be. The map never claimed every node is a pivotal-evidence node. Most upstream nodes are internal-decision nodes: they change what you build, not what you file. And the payer's boundary is my argument, not the rebuttal to it. If RWE can never supplant the RCT, then the place it pays is upstream, shaping the RCT you actually run. Focus beats boiling the ocean.
The map, stated once: the charted coastline is late, external control, submission, safety, market access, label expansion, all real, all where the landmark wins sit. The interior is upstream, disease sizing, natural history, feasibility, endpoint validation, all faintly drawn. Those nodes aren't empty; almost nobody stands there.
If you have one RWE pound left to spend this quarter, spend it inland. Model the feasibility, benchmark the comparator, prove the endpoint is ascertainable, all before the protocol locks. That's the one stretch of the map where a real-world input still changes the trial instead of merely describing it. Everywhere else, you're annotating a decision you already made.
At every stage of the lifecycle: sizing the disease at discovery, establishing the natural-history counterfactual, modelling feasibility and validating endpoints before the protocol locks, building an external control, assembling the submission, and running post-approval safety surveillance, market access and label expansion. Most teams only reach for it at submission and after approval, which leaves the upstream nodes unused.
Rarely, and only conditionally. RWE for safety, feasibility and disease sizing is common and rising, but RWE as pivotal efficacy evidence is uncommon and always earned against agency scrutiny. Its largest, least-banked value sits upstream, shaping the trial you actually run rather than replacing it.
If you're running or planning a single-arm trial, build the external comparator early, to the standard the agency will actually apply rather than the one you wish it would. Acceptance is not automatic: poorly designed external controls are regularly judged non-supportive, so design to the agency's bar from the start.
Several core FDA real-world data and real-world evidence guidances are final, but two that matter most for pivotal use, covering externally controlled trials and non-interventional studies, remain in draft. Map each RWE element in your package to the specific guidance that governs it, and flag anything standing on draft guidance for early agency discussion.
Get the monthly digest
The 5 things evidence leads need to know each month: regulatory moves, RWE developments and what they mean in practice. No pitch, one email a month.
[1] 21st Century Cures Act, §3022; codified at 21 U.S.C. §355g (enacted 13 December 2016). Cornell Legal Information Institute. https://www.law.cornell.edu/uscode/text/21/355g
[2] European Medicines Agency. Reflection paper on the use of real-world data in non-interventional studies to generate real-world evidence (EMA/99865/2025). 3 April 2025.
[3] Dagenais S, et al. (2022). Clin Pharmacol Ther. PMID: 34839524. https://pubmed.ncbi.nlm.nih.gov/34839524/ (NET prevalence, ALS A4V-SOD1 natural history, haemoglobin surrogate and Trial Pathfinder figures are secondary within this review, attributed to the primary sources it cites.)
[4] Griffith SD, et al. (2019). Adv Ther. PMID: 31140124. https://pubmed.ncbi.nlm.nih.gov/31140124/
[5] Arondekar B, et al. (2022). Clin Cancer Res. PMID: 34667027. https://pubmed.ncbi.nlm.nih.gov/34667027/
[6] Vaghela S, et al. (2024). Orphanet J Rare Dis. PMID: 38475874. https://pubmed.ncbi.nlm.nih.gov/38475874/
[7] Alipour-Haris G, et al. (2024). Clin Transl Sci. PMID: 39092896. https://pubmed.ncbi.nlm.nih.gov/39092896/
[8] Crisafulli S, et al. (2025). Drug Saf. PMID: 40223041. https://pubmed.ncbi.nlm.nih.gov/40223041/
[9] Boss J, et al. (2026). Int J Technol Assess Health Care. PMID: 41508412. https://pubmed.ncbi.nlm.nih.gov/41508412/
[10] Deng Y, et al. (2025). Ther Innov Regul Sci. PMID: 40468096. https://pubmed.ncbi.nlm.nih.gov/40468096/
[11] U.S. Food and Drug Administration. Framework for FDA's Real-World Evidence Program. December 2018.
[12] U.S. Food and Drug Administration RWD/RWE guidance suite. FINAL: "Considerations for the Use of Real-World Data and Real-World Evidence to Support Regulatory Decision-Making for Drug and Biological Products" (August 2023); "Submitting Documents Using Real-World Data and Real-World Evidence to FDA for Drug and Biological Products" (September 2022); "Assessing Electronic Health Records and Medical Claims Data to Support Regulatory Decision-Making" (July 2024); "Assessing Registries to Support Regulatory Decision-Making" (December 2023). DRAFT: "Externally Controlled Trials for Drug and Biological Products" (February 2023); "Real-World Data: Assessing Electronic Health Records and Medical Claims Data" / "Non-Interventional Studies" draft guidance (March 2024).
[13] ICH M14: General Principles on Plan, Design and Analysis of Pharmacoepidemiological Studies That Utilize Real-World Data for Safety Assessment. DRAFT, Step 2b, 21 May 2024.
[14] National Institute for Health and Care Excellence. NICE real-world evidence framework (ECD9). 23 June 2022.
[15] FDA Approval Summary: palbociclib (Ibrance) for male breast cancer. PMID: 31649043. https://pubmed.ncbi.nlm.nih.gov/31649043/
[16] Effectiveness of tacrolimus (Prograf) in lung transplantation, benchmarked against SRTR data. Transplantation. 2022;106(6):1233–1242. PMID: 34974456. PMC9128622. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9128622/
[17] ClinicalTrials.gov records: NCT00468832 (CINRG Duchenne natural history); NCT01093404 (TASTE); NCT02697916 (ADAPTABLE); NCT01551758 (Salford Lung Study, COPD); NCT01907087 (cerliponase alfa, CLN2); NCT03306277 (Zolgensma STR1VE); NCT06019637 (ARISER, Zolgensma PASS); NCT06785818 (tisagenlecleucel real-world registry, Korea). https://clinicaltrials.gov/study/NCT00468832