External Comparator Groups Derived from Real-world Data Used in Support of Regulatory Decision Making: Use Cases and Challenges

Real-world data (RWD) from electronic health records (EHRs) and administrative claims databases are used increasingly to generate real-world evidence (RWE). RWE is used to support clinical evidence packages for medicines that inform decision-makers. In this review of current issues in the use of RWD-derived external comparator groups to support regulatory filings, we assess a series of topics that generally apply across many disease indications. However, most of the examples and illustrations focus on the oncology clinical research setting. The topics include an overview of current uses of RWD in drug development, a discussion of regulatory filings using RWD-derived external comparators, a brief overview of guidance documents and white papers pertaining to external comparators, a summary of some limitations and methodological issues in the use of external comparator groups and finally, a look at the future of this area and recommendations.

decision making (FDA Framework for RWE), 1 and research initiatives such as the 2017 National Cancer Institute Cancer Moonshot [31-33, 36, 37].The changing regulatory landscape is not unique to the USA, as health authorities in other jurisdictions, such as the European Medicines Agency (EMA), the Pharmaceuticals and Medical Devices Agency (PDMA) in Japan, and Health Canada, have all issued recent statements regarding the developing role of RWD in drug development [38][39][40].
Changes in regulatory practice are not the only reason for the increased focus on externally derived control groups. The field of medicine has benefited from the advent and availability of next generation gene sequencing (NGS) platforms that have greatly enhanced oncology drug discovery and led to increased targeting of onco-genic mutations [41]. Many targeted oncology molecules are now following accelerated pathways such as FDA's Breakthrough Therapy Designation, with a commensurate acceleration in the drug development cycle that can advance experimental therapies from phase 1B directly into phase III trials. For example, sortorasib was recently granted accelerated approval for KRAS G12C-mutated locally advanced or metastatic nonsmall cell lung cancer (NSCLC) [42]. In addition to speeding the drug development cycle, externally derived control groups can support situations in which randomization may not be possible or ethical, including when effective treatments are not available (e.g., novel biomarker targets) or current treatment options are suboptimal. There are numerous examples in oncology drug development where the unmet medical need is so pronounced that palliative care would be the only alternative to an experimental treatment. In yet other cases, the populations have rare biomarkers that make patient enrollment challenging. Collectively, these factors have resulted in a growing number of early phase single-arm oncology trials and increasing attention to "hybrid' designs in later phase trials, which include randomized controls augmented with external controls.
The availability, completeness, and quality of RWD, especially EHR databases, has been steadily improving. These EHR databases can augment structured data fields with unstructured information gleaned from text fields in the medical charts. Also, the data are more contemporaneous, with very little lag between the time of a medical encounter and the data becoming available for analysis. There is also better characterization of important biomarkers as testing practices increase over time. Linkage of data sources such as EHRs with administrative claims databases provides additional granularity and completeness and fills in important missing gaps in the patients' clinical, treatment and demographic profiles. Finally, the overall quality of important endpoint and clinical outcomes like mortality and disease response has been critically assessed and validated in some EHRs [43][44][45].
The use of RWD for externally derived comparator groups raises important methodologic considerations. To address selection bias and other forms of study bias, statistical methods to control for bias and confounding have continued to evolve [46][47][48][49][50][51][52][53]. Developing methods include the use of target trial emulation principles to avoid various forms of bias, including selection bias, the use of summary confounder scores such as a propensity score (PS) and methods for assessing and controlling unmeasured confounding, as accomplished with negative controls and use of instrumental variables [46][47][48][49][50][51].
In this review of current issues in the use of RWD-derived external comparator groups to support regulatory filings, we assess a series of topics that generally apply across many disease indications. However, most of the examples and illustrations will focus specifically on the oncology clinical research setting. The topics included in the review are as follows: • An overview of current uses of RWD in drug development • Regulatory filings using RWD-derived external comparators • Guidance documents and white papers pertaining to external comparators • Limitations and methodological issues in the use of external comparator groups • A look at the future of this area and recommendations

Overview of Uses of RWD-derived External Comparators
Historical controls generally refer to the use of patient cohorts derived from previously conducted clinical trials that are repurposed for use in the assessment of treatment effects or adverse events observed in other clinical trials.
In addition to concerns about limited availability of critical variables, the use of historical control arms as comparators for single-arm trials have been criticized because the historical data may not reflect current standard of care. The increasing availability of more contemporary RWD enables concurrent comparisons, and accordingly, the nomenclature for these types of applications has evolved from historical controls to external, synthetic, or virtual controls/comparators [3,6,10,54,55].
Several uses exist for combining externally derived cohorts with clinical trial data. The corresponding point in a medicine's development lifecycle where an application is relevant is depicted in Fig. 1. These main applications are elaborated on in Table 1 and are discussed below.

Single-Arm Trial Settings
Disease burden and unmet medical need can be explored by examining early stage clinical development findings with the natural history of the disease using stand-alone externally derived cohorts [34,56]. The overarching purpose is to characterize unmet medical need for a given disease target population, which in turn provides an understanding of the potential benefits offered by an experimental treatment in the absence of a randomized control group [34,56]. This approach is often warranted owing to outdated literature or lack of publications focused on the specific target population of interest.

Single-Arm Trials: Direct Comparisons Using External Comparator Arms
When appropriate RWD are available they can be used to make direct comparisons to trial experimental arms. RWD-derived external comparators can mimic the overall health, demographic, and disease characteristics of the trial arm through application of a trial's inclusion and exclusion criteria ( Fig. 2A). However, these non-randomized comparisons are subject to potential biases. For example, the prevalence of a biomarker that defines the trial population could differ in the RWD control group, introducing selection bias and threatening comparability. To reduce selection bias, the RWD-derived controls should be sampled from a similar underlying population as the treatment arm. To compensate for potential bias and confounding between the trial arm and the externally derived comparator arm, appropriate adjustment methods (e.g., propensity score matching and weighting) and sensitivity analyses should be incorporated into the analyses [6,8].

Phase I/Ib
Despite growing interest and potential, relatively few studies have used external controls as a means of interpreting data from clinical trials. One of the first examples is a study by Gökbuget et al. comparing outcomes from a single-arm trial in relapsed/refractory acute lymphoblastic leukemia patients to an external comparator [57]. These researchers constructed an external control group using RWD to compare response and survival between the external comparator and clinical trial. The data were included in regulatory filings and played an important role in the FDA's accelerated approval of blinatumomab [58]. The efficacy of blinatumomab versus standard of care chemotherapy was further confirmed in a phase 3 randomized controlled trial two years later [59]. The concordance of the findings provided additional support for the use of external controls.
Tan and colleagues took a different approach than the blinatumomab example by extracting data from published single-arm trials in patients treated with crizotinib for anaplastic lymphoma kinase (ALK) positive metastatic non-small cell lung cancer (mNSCLC) [60] and comparing the aggregate findings with single-arm studies in NSCLC patients treated with ceritinib. The comparison demonstrated an advantage for crizotinib in progression-free survival (PFS) and overall survival (OS) [60]. Other research using external controls have also shown early promise [12,54,55], including studies drawing on aggregate findings from the clinical trial literature in acute myeloid leukemia and anaplastic lymphoma kinase-targeted NSCLC; each of these efforts demonstrated the promise of external controls for interpreting findings from single-arm trials [54,60,61]. A recent study drew on data from 11 advanced NSCLC (aNSCLC) randomized trials and substituted data from an oncology EHR database [8] for the controls arms from the clinical trials. In most cases, the researchers were able to replicate closely the hazard ratios for overall survival from the original randomized trials [8].

Accelerating Late Phase Trials: Hybrid Designs Using External Controls
Innovative study designs for late phase oncology trials are another opportunity for using external controls. In the past, the purpose of so-called adaptive trials was to channel more patients to the RCT experimental arm that was experiencing better outcomes [35]. Hybrid trials have a couple of different meanings by regulators [35]. There are hybrid trial designs that use EHR data to collect information on patients enrolled in trials and thereby reduce costs and timelines (see Zhu et al. 2020 for a comprehensive treatment of this design [62]). Another form of hybrid trial design, which shares similar objectives to both adaptive designs and the previously mentioned hybrid trials, are trials that augment randomized controls with controls from other trials or from RWD that share similar characteristics as those in the trial [1,26,62]. These hybrid designs expose a smaller proportion of randomized subjects in the clinical trial to a potentially suboptimal standard of care, and the external data (historical clinical trial and/or RWD) are used to supplement the randomized control. In some cases, Bayesian approaches are used to assess commensurability between the external controls and the trial patients over time [1,4,62]. The assessment of the commensurability between the trial data and the external cohort takes place incrementally over the course of the trial enrollment period and may involve techniques that select controls from the external data based on the assessment of commensurability with respect to the trial outcome (e.g., the greater the commensurability, the more external patients selected to augment the trial patients). Figure 2B shows a simplified version of the hybrid trial design using a post hoc "all or none" selection of external controls approach in a fictional trial where a 3:1 randomization scheme is used. The purpose of these approaches is to accelerate trial enrollment, expose fewer patients to suboptimal treatments, and ultimately bring innovative potentially lifesaving or life-extending medicines to patients sooner.

Indirect Comparative Effectiveness
Indirect treatment comparison in the pre-approval and postmarketing settings allow for comparisons between experimental treatments or newly marketed drugs respectively, with novel or new standard of care marketed treatments.
Using external RWD cohorts standardized to trial eligibility criteria, treatment effects are characterized between a trial experimental treatment and some standard of care treatment in the RWD [60,61]. The RWD cohort is then re-standardized to a published trial that compares the same standard of care with another newer novel marketed therapy [63]. Additional adjustment can be achieved through weighting on aggregate level baseline characteristics [63]. The analysis involving the published trial data can be accomplished using software that converts digital Kaplan-Meier (KM) curves using a pixel-based conversion approach that reconstructs the published KM curve with the RWD external comparator in the same KM plot [63,64]. The same technique may be used to extract a novel treatment from the published literature and plot a KM graph comparing the novel treatment with an experimental treatment from a trial.

Post-marketing Safety Studies and Long-term Safety Follow-up
Post-marketing safety studies containing a single exposed cohort is another setting where the use of externally derived cohorts can aid in the interpretation of study findings [19]. For example, in prospective studies examining rare populations where an internal comparator is infeasible due to enrollment challenges (e.g., idiopathic pulmonary fibrosis), an external comparator arm derived from a large administrative claims database can be beneficial for evaluating important safety endpoints. Often pregnancy safety exposure registries will capture data only on exposed patients, making interpretation of potential safety signals challenging in the absence of an unexposed control group. These studies may rely on the published data to interpret safety signals, which may be outdated and may not reflect the same patient characteristics found in the registry. Finally, long-term safety studies could be conducted using data from RWD-derived patient cohort as an extension to the trial observation period allowing for adequate follow-up of important safety endpoints [6].

Regulatory Filings Using RWD-derived External Comparators
Regulatory reviews of hematology and oncology new drug applications (NDAs) and labeling extensions submitted by drug developers using data from single-arm trials is increasing [7,16,24,[65][66][67] (Fig. 3). Despite the growth in single-arm study approvals, RWD have been included in very few evidence packages. This trend could change in response to the December 2016 enactment of the 21 st Century Cures Act and the subsequent initiation of the NCIs Cancer Moonshot program in 2017. Bolislis et al. (2020) reviewed NDAs and labeling extensions using RWD submitted to the EMA, FDA, PMDA, and Health Canada over the past 20 years and found only 27 cases where RWD was used, primarily in the oncology disease area [16]. In addition to the blinatumomab example mentioned above, there are five other examples of favorable FDA decisions that considered RWD. These include label expansions for blinatumomab, avelumab approval for metastatic Merkel cell carcinoma, and axicabtagene ciloleucel approval for relapsed or refractory large B-cell lymphoma [16,68]. Although not an example of an RWD external comparator analysis, palbociclib received an approval in metastatic male breast cancer based in part on evidence from RWD [69].
There are also examples where drug developers have been unsuccessful using RWD-derived control arms as part of their FDA filing. Despite receiving Breakthrough Therapy Designation for relapsed or refractory multiple myeloma (RRMM), in the selinexor submission, the regulators commented on study design issues, differences in the EHR population compared to the trial patients, and methodologic issues, which led them to conclude that the RWD "is not adequate to provide context or comparison for the overall survival observed" in the clinical trial patients [70]. Additionally, the regulators expressed concerns about the RWD study not being pre-specified, as the study protocol was not submitted a priori to the FDA [70]. In 2019, the sponsor of tazemetostat, indicated in epithelioid sarcoma patients, had their RWD evidence disregarded by the FDA due to lack of pre-specification of the study protocol and concerns about the study design and methods [71]. Again in 2019, the FDA did not consider the submitted RWD-derived control arms for the filings for both erdafitinib, a fibroblast growth factor receptor positive (FGFR +)-targeted therapy for patients with advanced or metastatic urothelial cancer, and entrectinib, for ROS1 + metastatic non-small cell lung cancer. The FDA cited concerns related more to the generalizability of the EHR cohorts than to concerns related directly to their methodologies [72,73].
In 2020, an external comparator study was submitted to the FDA (RE-MIND: NCT04150328) to support the NDA for tafasitamab indicated in patients with relapsed or refractory diffuse large B-cell lymphoma (DLBCL) [74]. This analysis aimed to provide context for interpreting the efficacy findings observed in the singlearm L-MIND pivotal trial for patients with DLBCL. The primary objective of the study was to isolate the contribution of tafasitamab to efficacy of a tafasitamab    [74]. In the FDA decision, the reviewers noted the following concerns: The validity of the study is compromised by several limitations in the study design. Bias resulting from key differences in patient selection and unequal distribution of important measured and unmeasured prognostic indicators between treatment arms are likely to favor survival for the L-MIND patients. Most importantly, given important differences in the patient populations included in the L-MIND trial and RE-MIND study, primarily as a result of selection bias, this study does not provide sufficient evidence to isolate the contribution of tafasitamab to efficacy of tafasitamab+LEN combination therapy for DLBCL" (Page 3) [74] This survey of reviews conducted by the FDA indicates that regarding the use of RWD-derived external comparators for contextualizing single-arm trials, these are at best early days in terms of acceptance. Concerns relate to selection bias, generalizability of RWD, and the resulting inherent differences in baseline covariates between RWD comparators and trial patients, as well as the capture and completeness of important prognostic factors in RWD.

Best Practices: Guidance Documents and White Papers
Given the growing interest in external controls, there are several recent guidance documents, best practices, and white papers in circulation devoted to the subject. FDA's Framework for Real-World Evidence Program discusses the role of external controls in the context of single-arm trials using either historical clinical trial data or RWD: Collection of RWD on patients currently receiving other treatments, together with statistical methods, such as propensity scoring, could improve the quality of the external control data that are used when randomization may not be feasible or ethical, provided there is adequate detail to capture relevant covariates. (page 20) [33] The Framework suggests limitations such as a lack of standardized diagnostic criteria and study endpoints, and general concern with the comparability of RWD patient populations with trial patients [33]. Other regulatory agencies, including EMA, PMDA, and Health Canada, have commented on the role of RWD in regulatory filings [38][39][40].
Many recent commentaries discuss the use of RWDderived external controls and their use in drug development [5,6,9,10,15,17,19,22]. Common recommendations for when to use RWD include single-arm trials where randomization is infeasible or unethical, label expansions, long-term follow-up, and augmenting randomized control groups in hybrid designs. An additional rationale relates to drugs that fill high unmet medical need where filings with single-arm studies receive accelerated approval (e.g., see sortorasib example above) and are followed by randomized phase 3 study as part of a post-marketing requirement.

Methodological and Other Considerations
This section highlights analytic issues (e.g., bias and confounding) that may arise when external RWD cohorts are combined with clinical trial data as well as some general considerations (see Table 2).
FDA and other regulators have repeatedly noted concern regarding how well RWD populations reflect clinical trial patients in terms of clinical and demographic characteristics and potentially unobserved prognostic factors. Although these differences can be controlled using statistical methods, this ability depends on data, especially prognostic factors, being available in the RWD. Additionally, differences in biomarker testing practices between RWD and trials can complicate studies specific to certain genetic mutations or protein expression. In the real-world setting patients may be tested late in the disease process, whereas patients enrolled in trials will be tested at the outset of the study. This difference in timing can affect trial results especially for those trials which are conducted in patients treated in the first line.
The use of concurrent controls from the RWD can help address changes to standard of care treatments that are common in oncology. Other issues in using RWD include endpoints like treatment response and disease progression that can differ in how they are defined and collected in clinical trials versus RWD [75]. Mortality, although well captured in some inpatient EHRs, is still less complete than found in clinical trials [43,44] and missing death data could lead to underestimation of mortality in the control arms [44]. In addition, many EHR databases only include US patients, whereas trial populations are often global and disease prognosis can vary by region.
Trials conducted in later lines of treatment can complicate selection of an appropriate index date or "time 0" for RWD analyses. A biomarker test date may occur after the initiation of treatment, introducing an "immortal" period Issues with comparability of RWD and trial patients and potential impact on study internal validity Are the RWD contemporaneous with trial period?
Changes in the standard of care can greatly affect analyses with outdated standard of care often causing bias in favor of the experimental treatment US-based EHRs versus global trials Regional differences in prognosis related to treatments and differences in health care delivery may threaten the external validity of the study findings Biomarker testing practices There may be differences between RWD patients and trial patients with respect to specific mutations or protein expression in terms of when the testing took place. Trial patients are tested at the outset of the study whereas costly NGS testing may take place very late in the disease process in patients in the real-world setting, often as a last resort Information bias Comparability of RWD endpoints Comparability of study endpoints: difference in important endpoints like disease progression between RWD and trials can introduce bias Efforts should be made to validate endpoint like mortality against known gold standard to understand the potential for underestimation of mortality in the RWD patients Data ascertainment Differences in timing of data collection and data measurement: Differences in biomarker testing practices with testing often conducted in later lines in RWD versus early on in trials; ECOG not routinely collected in RWD; serial measurement of lab data not as likely in RWD Design and analysis issues Handling and assessment of missing data Missing mortality in the external comparator may result in an overestimate of survival in the external comparator group introducing a bias in favor of the standard of care group Selection of comparator lines and time origin Selection of comparator lines and time origins may introduce selection bias and immortal time periods. Care must be taken so the selection of comparators and the start of follow-up doesn't require knowledge of things that occur in the future Bias analyses pre-specified Pre-specifying probable study challenges will enhance study credibility (e.g., handling of missing data and data related limitations that require certain assumptions)  [76,77]. Figure 4A and B illustrates this point in studies examining first-and second-line-treated patients respectively. There are multiple methods available to address possible shortcomings of RWD, including negative controls, instrumental variables, propensity scores (PS), and highdimensional PS [13,47,48,50,51,78]. Other methods such as inverse probability of censoring and time to censoring for handling potential informative censoring bias stemming from missing outcomes like mortality can improve the integrity of the analysis [52]. Quantitative bias analysis can be used to assess the robustness of findings and assumptions [53]. Approaches such as study restriction by line of treatment, stratification or statistical adjustment by line of treatment can also be used. An assessment of important confounders not captured in RWD may point to the need for additional data abstraction of these potential confounders. Many of these issues can be addressed through thoughtful study design and analysis.
Transparency about data limitations, comparability of treatment groups and other potential sources of bias is crucial. In oncology and other diseases areas with high mortality or morbidity, regulators must make benefit/risk determinations and a common understanding of what can, and cannot, be inferred from the RWD comparisons serves regulators and sponsors, but most of all patients.

The Future of External Comparators
Changes in the regulatory landscape have led to an increased focus on the use of RWD as external comparators in the clinical evidence package used in regulatory drug submissions in oncology. To date, best practices are still being established and debated among the scientific community with many groups such as the FDA, Drug Information Association, Friends of Cancer Research, American Society for Clinical Oncology, and the International Society for Pharmacoepidemiology currently working on guidance around the use of external comparators as part of the clinical evidence package [6,17,25,26,30,33,78]. Adoption of these approaches will hinge on continued dialogue and scientific exchange with a view to producing credible evidence to support drug development efforts. What does the future hold? Inevitably, there will be ever greater availability of RWD data sources in oncology with more granularity and improved quality, including growing capture of NGS and other biomarker test results. Methods to account for bias and confounding will continue to evolve. One of the major challenges will be to synthesize the empirical findings into a coherent model that captures the more salient learnings from a wide spectrum of researchers. Our ability to do this in a thoughtful and collaborative way across the various stakeholders will ultimately decide the role that RWD will play in drug development.
When viewed from the various stakeholder perspectives, whether patients, drug developers, regulatory agencies, or payers, there is a common shared interest in accelerating oncology drug development. For patients diagnosed with cancer, the prospect of novel treatments becoming available that may be able to help extend their life is obviously crucial. Biopharmaceutical companies have as their core mission the development of effective and safe treatments to combat lifethreatening disease. Single-arm studies can accelerate the approval process in settings with a high unmet medical need and get medicines to patients faster. Using RWD-derived control arms to evaluate efficacy can also inform better decision making. Regulators also have as their mission to bring safe and effective medicines to patients in a timely fashion. Finally, payers would benefit from the availability of more suitable targeted treatments with enhanced prognoses that would in the long run justify the cost of newly developed treatments. With all the apparent advantages that RWE can potentially provide to the traditional drug development paradigm come myriad concerns related to poorly designed RWD/clinical trials or RWD that is used in settings where it is not appropriate or warranted. These lingering concerns makes it even more important to focus on advancing the scientific knowledge base pertaining to the use of RWD in clinical drug development.