FDA Approvals Based on Natural-History and External Controls: What Reviewers Accepted

FDA approvals based on natural-history and external controls: what reviewers accepted and why

FDA has accepted a natural-history or other external control as the primary comparison for effectiveness in a small group of rare-disease approvals. Examples from CDER review records include Kanuma (sebelipase alfa), Strensiq (asfotase alfa), Nulibry (fosdenopterin), Zokinvy (lonafarnib), Brineura (cerliponase alfa), Sohonos (palovarotene) and Iwilfin (eflornithine). The list is not exhaustive: CBER gene-therapy approvals such as Zolgensma, Skysona and Lenmeldy were also compared with natural-history data and are not covered here. In other programs, external controls only supplemented randomized trials (Voxzogo, Yuviwel, Fabrazyme), or reviewers rejected them outright (Exondys 51, Forzinity).

The reviews give consistent reasons for acceptance. The untreated disease course was predictable and severe. The endpoint was objective, often mortality. Patient-level control data allowed matching on the prognostic factors that mattered, such as genotype, age and baseline severity. The effect was large, and it held up across sensitivity analyses. When these features were missing, FDA treated the external comparison as uninterpretable. That happened most often with effort-dependent endpoints assessed open-label, such as the 6-minute walk test.

The regulatory standard reviewers applied

FDA recognizes historical controls under 21 CFR 314.126 and considers them a subset of external controls. Its 2023 draft guidance, Considerations for the Design and Conduct of Externally Controlled Trials for Drug and Biological Products, says external controls can serve as the adequate and well-controlled investigation for substantial evidence when appropriate, subject to case-by-case assessment ⁴⁰²⁷. ICH E10 limits external controls to unusual circumstances. It says they are most persuasive when the disease course is well documented and highly predictable, the endpoint is objective, and the outcome in treated patients is markedly different from the control, with a high level of statistical significance ²²²³. The draft guidance adds more specific expectations:

• Comparable populations. The control should resemble the treated group on demographics, disease severity and duration, diagnostic criteria, prior and concomitant therapy, and the start of follow-up ²²²⁷³⁵.
• Consistent index date. "Time zero" should be defined the same way in both arms to avoid immortal-time bias ²⁸.
• Consistent outcome assessment. Outcomes should be defined and assessed the same way in both arms, and blinded where feasible ²⁸.
• Prespecification. The control source and analysis should be chosen before the trial, not retrospectively after a single-arm study ³².
• Suitable data. The natural-history guidance points out that poorly defined or unreliable outcome assessments in external-control data are a particular problem for regulatory use ²⁴.

The review packages below show how these principles played out in practice.

Summary of key programs

Programs where the external control was the primary comparison

Kanuma (sebelipase alfa): the textbook case

For infants with Wolman disease, the single-arm study LAL-CL03 was compared with LAL-1-NH01, a retrospective chart-review natural history. Of the 35 LAL-1-NH01 patients, 21 were adjudicated appropriate for comparison ⁹²³⁹. Six of nine treated infants (67%) survived to 12 months, compared with none of the 21 historical controls. All controls died before 8 months, at a median age of 3.5 months. The hazard ratio was 0.141 (95% CI 0.040 to 0.496) ²⁴⁰.

FDA accepted the control for these reasons ^9240243:

• The disease was clinically homogeneous and predictably fatal, with death expected within the first 6 months of life.
• Baseline characteristics were acceptably similar between cohorts.
• The endpoint, survival, was objective and had been recommended by the division.
• The difference was large and statistically significant.
• The Deputy Division Director called the natural-history study properly conducted.

Reviewers still noted the risk from unmeasured confounding and selection bias in a retrospective design, made worse by the very small sample ²⁴². They also noted that Kaplan-Meier hazard-ratio methods fit balanced concurrent cohorts better than a nonconcurrent historical comparison ²⁴⁰.

Strensiq (asfotase alfa): strong for infantile HPP, weaker for juvenile HPP

In perinatal/infantile-onset hypophosphatasia, pooled single-arm studies were compared with the retrospective natural history ENB-011-10 ⁹⁰. At last contact, 91.2% of treated patients were alive versus 27.1% of historical controls ⁵⁶. Ventilator-free survival was 96% versus 31% ¹⁰².

FDA's reasons for accepting the comparison were ¹³⁸⁹⁵⁶:

• Infantile disease was phenotypically homogeneous and expected to be fatal in the first year of life.
• Baseline characteristics were similar.
• The endpoints (survival and ventilator-free survival) were objective and had been recommended by FDA.
• The differences were large.
• Sensitivity analyses confirmed the finding, including worst-case censoring, Nelson-Aalen estimation and an analysis restricted to patients on the intended dose.

FDA still classified the hypothesis tests as exploratory, because endpoints and historical-control comparisons were set during study conduct ⁸⁶⁹⁰.

The juvenile-onset comparison was weaker. It used the chart-abstraction study ALX-HPP-502, with 8 treated patients versus 32 controls ⁹³⁸⁷. FDA judged juvenile HPP too heterogeneous and less fulminant for a natural-history comparison to be compelling ⁴. Radiograph intervals were unequal between arms ⁸⁸, and the gait instrument was post hoc and not validated in HPP ⁹⁹¹⁰². The statistical reviewer believed a randomized trial was feasible ⁴⁸⁷. Juvenile efficacy was therefore accepted on the totality of descriptive evidence, with growth as the lead parameter, rather than on the external comparison alone ⁹⁰⁹¹.

Nulibry (fosdenopterin): genotype matching and a single-study framework

FDA described Nulibry's evidence as one adequate and well-controlled investigation plus confirmatory evidence ¹³⁹¹⁵⁴. The investigation compared 13 treated patients with 18 genotype-matched untreated controls. Deaths occurred in 2 of 13 treated patients (15%) versus 12 of 18 controls (67%), giving a hazard ratio of 0.18 (95% CI 0.04 to 0.72) ¹³⁵¹⁴⁹. Unadjusted hazard ratios ranged from 0.17 to 0.24 across the full untreated cohort, the genotype-matched set, weighted analyses and sex-adjusted analyses ¹³⁸¹⁴².

The review explained why the control was adequate:

• Genotype matching. Genotype predicts phenotype in this disease, and FDA concluded that matching on it adequately addressed selection bias ^13351146.
• Sensitivity analyses. The conclusion did not change when FDA restricted to exact genotype matches, excluded a pair with discordant symptom-onset age, excluded a cross-region match (United States vs Malaysia), or limited the analysis to contemporaneous births ^13314351.
• Detection bias ruled out. All prenatally diagnosed treated patients developed symptoms within the first month of life. The same genetic testing was available to historical controls ⁵¹¹³⁹.
• Confirmatory evidence. Urinary S-sulfocysteine reduction, exposure-response data and survival in a mouse model supplied the confirmatory evidence ¹⁴⁰¹⁵⁴.

Acknowledged weaknesses included baseline imbalances in sex, birth era and race, a small sample, and a statistical analysis plan written after mortality outcomes were known ^145146140.

Zokinvy (lonafarnib): FDA re-engineered the matching

Survival in the single-arm ProLon1 and ProLon2 studies was compared retrospectively with untreated patients, drawn mainly from the Progeria Research Foundation International Progeria Registry ²²⁷²²⁵. FDA found the sponsor's original random matching inadequate for two reasons. First, it produced more nonclassic-variant patients among controls (11%) than among treated patients (3%). Nonclassic disease has worse survival, so the imbalance favored the drug ⁵⁰. Second, the control set changed depending on how the random selection was run ⁵⁰.

FDA made three changes:

• It required matching on variant status, sex and continent of residence, using a deterministic fixed-50th-percentile selection ²¹³.
• It required matching censoring of untreated controls, to correct differential censoring ²¹⁷.
• It treated alternative matching schemes as supportive analyses ²¹⁴.

In the main analysis (62 treated vs 62 matched controls), the mortality hazard ratio through last follow-up was 0.40 (95% CI 0.21 to 0.77), with a mean survival gain of 2.5 years ²²⁴.

The review cited these factors as making the control adequate:

• Controls were contemporaneous, which made time-period effects less likely ²²⁶.
• There was no evidence that changes in supportive care had affected HGPS life expectancy ²¹⁶.
• Global recruitment and trial-access support reduced selection bias ²²¹²²³.
• Mortality is an objective endpoint, which offset the post hoc analysis plan ²²⁶.
• The Office of Surveillance and Epidemiology audited vital-status data and found high concordance ²¹⁶.

FDA granted regular approval on November 20, 2020 ²⁸³²⁸⁴. Residual concerns included few deaths, wide confidence intervals and no matching on birth year ²²³.

Brineura (cerliponase alfa): acceptable only after conservative restrictions

The single-arm studies 190-201/202 were compared with Study 190-901, an untreated cohort from the DEM-CHILD registry built largely from retrospective chart and interview data ¹⁸⁴⁵⁸. Several review findings shaped the analysis:

• The CLN2 rating scales used in the two sources had different anchors and training.
• A video comparability exercise suggested that Language-domain ratings were biased toward treatment.
• FDA therefore limited the primary endpoint to the Motor domain ¹⁷⁶¹⁷⁸.
• It required a sustained 2-category decline to reduce the influence of measurement error ¹⁷⁶¹⁷⁹.

The final matched analysis paired treated and control patients on age (within 3 months), genotype and baseline Motor score. It showed a 59-percentage-point responder difference at 96 weeks (95% CI 24% to 83%) in 17 pairs ¹⁹⁰¹⁸⁸. Time-to-decline, ordinal and logistic analyses were concordant ¹⁸⁸.

FDA also asked for conservative handling of the historical data. The sponsor compared prospective-only data with the full cohort and assumed control decline occurred on the day it was documented ¹⁷². The 48- and 72-week data did not establish efficacy ¹⁷⁸. FDA granted traditional approval on April 27, 2017, with postmarketing safety and immunogenicity requirements ^313320316319.

Sohonos (palovarotene): same-site prospective natural history

The pivotal Study 301 was single-arm. Its control was the sponsor's prospective FOP natural-history study, run at the same centers with similar eligibility criteria, restricted to the R206H genotype, and assessed with blinded whole-body CT ^24825225368. FDA listed four conditions for relying on an external control in this program ²⁵²:

1. The natural history is well defined.
2. The control population is very similar to the treated population.
3. Concomitant treatments that affect the endpoint are not materially different.
4. The results compellingly show a change from established disease progression.

FDA found the sponsor's prespecified Bayesian model biased against the drug, because Study 301 had more frequent assessments. It accepted post hoc linear mixed-effects analyses instead, which estimated 9.4 versus 20.3 cm³/year of new heterotopic ossification, a difference of −10.9 cm³/year (95% CI −21.2 to −0.6) ^248257260. Propensity weighting, 1:1 nearest-neighbor matching and a within-subject analysis of 39 patients who moved from the natural-history study into Study 301 all supported the result ²⁴⁹.

Reviewers named reliance on post hoc analyses and an external control as the main evidentiary weaknesses ²⁴⁸²⁵⁷. They also flagged baseline imbalances, possible selection effects in who moved into the trial, and the uncertain meaning of negative changes in HO volume ^18255251252. Approval on August 16, 2023 followed a Complete Response Letter issued December 23, 2022 ²⁶⁶²⁶⁸.

Iwilfin (eflornithine): a prior trial arm as the control

For Iwilfin, the external control came from patient-level data in the Children's Oncology Group trial ANBL0032, not from a registry. FDA had recommended patient-level data rather than a historical rate ⁸⁰. Treated patients were propensity matched 1:3 to controls on 10 covariates, with exact matching on MYCN status ⁷⁰⁸⁰. Most treated patients (94%) had received their immunotherapy on ANBL0032 itself ²³⁴. The event-free survival hazard ratio was 0.48 and the overall survival hazard ratio was 0.32 ²³⁰⁷⁰. FDA said no single analysis would be primary. It weighed the primary, sensitivity and supportive analyses together ⁸⁰.

The review concluded the study was adequate and well controlled for these reasons ^234231235:

• The disease has a well-defined natural history and a poor prognosis.
• The estimated effect was large.
• Sensitivity analyses gave consistent results.
• Follow-up was sufficient, with few patients lost.
• Inspections found no data-integrity problems.

The review also named limitations:

• The external-control design was built after Study 3(b) results were published in 2018 ⁸⁰²³⁵.
• Immortal-time bias required a conservative 123-day sensitivity analysis ²³².
• Blinded central review was possible for treated patients but not for controls ²³⁰.
• Advisory committee members were not unanimous about relying on the external-control trial ²³⁶.

Programs where the external control was supportive only

Voxzogo (vosoritide). The AchNH registry was used to show long-term durability of the growth effect, not to establish efficacy, because long-term placebo data did not exist ^19167171. FDA asked for re-matching on baseline height and annualized growth velocity. This reduced the 5-year cross-sectional estimate from 9.08 cm to 8.15 cm ¹⁹¹⁶⁷.

In the S-002 supplement for younger children, the external controls overestimated the randomized effect. The Year-1 height Z-score difference was 0.25 against randomized placebo, but 0.51 against AchNH. FDA attributed part of the gap to inconsistent length and height measurement in the registry ⁶⁷. It is a clear example of an external control inflating the effect when a randomized comparison is available to check it.

Yuviwel (navepegritide). Effectiveness was established in a randomized, placebo-controlled trial ²⁴⁷. A matched natural-history comparison supported the long-term extension data. The statistical reviewer verified the matching, but 14 of 54 treated participants could not be matched, which FDA said may limit generalizability ²⁴⁴.

Fabrazyme (agalsidase beta). The 2021 supplement sought traditional approval. Its external comparison matched treated registry patients to a retrospective untreated natural-history cohort ²⁰⁵¹⁹⁸. FDA's epidemiology reviewers rated the study low-level evidence (Level II-3) ¹⁹⁵. They wrote that selection and information bias could completely explain the observed differences ¹⁹⁴. Their reasoning was that Fabry disease is heterogeneous, benefit is delayed, and the effect was not large, so the features that make historical controls credible "might not apply" ². The study was accepted only as supplementary evidence alongside the randomized trial AGAL-008-00 ¹⁹⁴.

Programs where FDA rejected the external control

Exondys 51 (eteplirsen). The sponsor proposed approval mainly on a post hoc comparison of 12 treated patients with 13 registry controls, using the 6-minute walk test ⁵⁴¹²⁶. FDA cited ICH E10's criteria and concluded Study 201/202 was "not a well-designed historically controlled trial" ¹²⁶⁶². Its concerns were:

• The open-label 6MWT is effort-dependent and can be influenced by motivation and coaching ¹²⁶⁴⁴.
• Controls started steroids later and had lower baseline NSAA scores ⁴⁴.
• Baseline 6MWT was not a matching criterion ⁵⁴.
• The registries were selected after more than three years of treated data were available ⁴⁴.
• The natural history of loss of ambulation was too variable to show that any treated patient had exceeded the expected range ⁵¹³¹.

FDA granted accelerated approval under 21 CFR 314.510 based on dystrophin expression ³⁰¹³⁰⁵. It required a confirmatory randomized trial with the North Star Ambulatory Assessment as the primary endpoint ³⁰⁴.

Forzinity (elamipretide). After the randomized SPIBA-201 Part 1 trial failed both primary endpoints, the sponsor relied on an external comparison, SPIBA-001 ⁸⁴. FDA called SPIBA-001 not an adequate and well-controlled study and its results "uninterpretable" ⁸⁴¹⁵. Its reasons were:

• Only 19 of 79 natural-history patients qualified ¹⁰⁴.
• Endpoint values at Weeks 64 and 76 were 100% imputed in both cohorts ¹⁰⁵.
• The propensity model used three covariates, and its code was dated before the analysis plan was finalized ^104105107.
• Treated patients' expectation of benefit on an effort-dependent endpoint could not be corrected with an external control ¹²².

After a May 15, 2025 Complete Response Letter, FDA granted accelerated approval on September 19, 2025. The basis was knee-extensor strength measured by handheld dynamometry, with a randomized placebo-controlled confirmatory trial required ^285287290289.

What made an external control adequate: patterns across the reviews

1. Predictable, severe natural history. Wolman disease, infantile HPP and MoCD Type A were described as homogeneous and rapidly fatal ⁹¹³. FDA made the opposite point for juvenile HPP, Fabry disease and DMD ambulation, where heterogeneity weakened the comparison ⁴²⁵.
2. Objective endpoints assessed consistently. Mortality carried the most weight: Kanuma, Strensiq, Nulibry and Zokinvy all used it, and FDA said for Zokinvy that it offset a post hoc plan ²²⁶. Blinded imaging worked for Sohonos ⁶⁸. Open-label 6MWT failed for Exondys 51 and Forzinity ¹²⁶¹²². Differences in rating scales forced Brineura's endpoint to be narrowed ¹⁷⁶.
3. Patient-level data matched on the prognostic factors that drive outcome. These were genotype for Nulibry ¹³³, variant status for Zokinvy ²¹³, MYCN status and immunotherapy exposure for Iwilfin ⁸⁰²³⁴, and baseline severity for Brineura and Voxzogo ¹⁹⁰¹⁶⁷. Missing key covariates weakened Fabrazyme and Forzinity ⁴⁹¹⁰⁴.
4. Contemporaneity and shared setting. Same-site enrollment for Sohonos ²⁵², a prior trial arm for Iwilfin ²³⁴, and contemporaneous controls for Zokinvy ²²⁶ all reduced concerns about time and care-setting effects.
5. Large effects that hold up in sensitivity analyses. Every accepted program showed consistency across alternative matching, censoring and model choices ^{89133214234188}.
6. Willingness to accept post hoc plans, within limits. Nulibry, Zokinvy and Sohonos were accepted despite post hoc analysis plans ^140216257. In each case FDA pointed to an objective endpoint, a large effect or confirmatory evidence as compensation. For Exondys 51, it judged the post hoc changes unreasonable even for generating hypotheses ⁶².

For sponsors, these reviews show that the external-control data source, the matching variables and the endpoint all need to hold up under scrutiny. FDA's inherent reservation about nonrandomized comparisons appears in every review, including the ones it accepted ⁹⁷⁰²⁵¹. Those approvals came from rare, severe diseases with objective endpoints and large effects, backed by the kind of control data these reviews describe.