Investigational Use of Real‐World Data as a Hybrid Control in Pancreatic Ductal Adenocarcinoma From the Randomized Phase Ib/II MORPHEUS Trial

Abstract

Enrolling adequate numbers of patients into the control arm of randomized controlled trials (RCTs) often presents barriers. There is interest in leveraging real‐world data (RWD) from electronic health records (EHRs) to construct external control (EC) arms to supplement RCT control arms and form hybrid control (HC) arms. This investigation showed the use of an HC arm in second‐line metastatic pancreatic ductal adenocarcinoma (PDAC). The RCT experimental arm (atezolizumab + PEGylated recombinant human hyaluronidase (Atezo + PEGPH20)) was compared with an HC arm consisting of patients treated with modified FOLFOX6 or gemcitabine/nab‐paclitaxel from the MORPHEUS PDAC internal control arm supplemented with data from a nationwide EHR‐derived de‐identified database as the EC arm. The EC arm was constructed by applying key inclusion/exclusion criteria from the MORPHEUS PDAC trial to patients from the real‐world cohort. Baseline variables were balanced using propensity score matching and covariate adjustment. Three analysis approaches—Cox model with pooled‐control data, Cox model with control arm‐specific frailty, and Bayesian analysis using a commensurate prior—were assessed. Overall survival was similar between the treatment arms. The direction and magnitude of hazard ratios (HRs) from the multiple HC analyses (HRs ranged from 1.02 to 1.06) were comparable with the reported trial HR (HR 0.91; 95% CI: 0.56, 1.49). This analysis demonstrates the feasibility and applicability of leveraging RWD in clinical trial design to supplement clinical trial control arms.

Study Highlights.

• WHAT IS THE CURRENT KNOWLEDGE ON THE TOPIC?

Randomized controlled trials (RCTs) are the gold standard for clinical trials, but difficulty in patient recruitment, control arm randomization, and cost limitations remain in some situations. Large real‐world databases of de‐identified patient data can be used to supplement trials.

• WHAT QUESTION DID THIS STUDY ADDRESS?

This study addresses the feasibility of supplementing an RCT internal control (IC) arm with external real‐world data (RWD) derived from electronic health records and informs how RWD data could be leveraged to derive treatment effects and precision estimates to aid trial decision‐making.

• WHAT DOES THIS STUDY ADD TO OUR KNOWLEDGE?

This study offers an actual application of a hybrid control (HC) arm from an ongoing clinical trial and was not based on hypothetical assumptions/simulations. Results utilizing the HC arm were equivalent to results obtained only using the RCT IC arm; however, the precision of the trial results was increased.

• HOW MIGHT THIS CHANGE CLINICAL PHARMACOLOGY OR TRANSLATIONAL SCIENCE?

These methods introduce approaches that may be generalized to future hybrid control analyses using RWD for more efficient trial designs.

In randomized controlled trials (RCTs), patients are randomly assigned to either an experimental or concurrent control arm, ideally in a double‐blinded fashion, to evaluate the efficacy and safety of new treatments. Although RCTs represent the gold standard for regulatory approval and reimbursement of new treatments, difficulties may be encountered when randomizing patients to the control arm in certain clinical settings. Targeted trials, including biomarker‐specific populations or trials with orphan drugs, are conducted with relatively few patients, and enrolling adequate numbers of patients often represents a challenge. These challenges are reflected in trials for rare populations or orphan drugs; the majority of these are designed as single‐arm trials. ¹ Additionally, enrolling patients into the control arm of confirmatory trials can be difficult when the new treatment is already available to patients due to accelerated approval. ² Similar enrollment challenges are faced by early‐phase trials (Phase Ib/II), due to the substantial cost of conducting well‐powered trials in this setting. Further, the fear of being randomly assigned to a control arm represents one of the top reasons for lower patient participation in clinical trials. ³

With these limitations associated with enrolling patients into the control arms of RCTs and given the increasing availability of high‐quality external data sources, interest is growing in leveraging these data. ⁴ External sources may include patient‐level data from completed RCTs or real‐world data (RWD), including electronic health records (EHRs), patient registries, and administrative insurance claims data. External data can be used as the sole control arm (external control (EC) arm trial) for a single‐arm trial or to supplement the concurrent internal control (IC) arm of an RCT (hybrid control (HC) arm trial). In single‐arm trials with an EC arm, it is not possible to determine what, if any, differences exist between the historical control population and a concurrent control population. ⁵ However, in HC arm designs, while fewer patients are randomized to the IC arm of the trial, the EC information and the IC data can be compared. Using this HC design, the relevancy of the EC arm can be evaluated while maintaining the scientific advantages of an RCT. ⁶ Hence, when concurrent controls are not precluded by ethical or practical concerns, HC‐arm trials are preferred over EC‐arm trials. ⁷ While the methodological applications to borrow information from an HC arm are not new, almost all publications have used historical trial data as EC data and/or focused on methodological aspects. ⁸ , ⁹ , ¹⁰ By comparison, limited information is available on the application of RWD as external data in HC arm trials.

Here, we report an analysis to assess how well an HC arm, consisting of data from patients enrolled in the IC arm from an RCT combined with RWD patients, can approximate the results of the trial IC arm and treatment‐effect estimates. MORPHEUS is a clinical trial platform comprising a suite of global, open‐label, randomized, umbrella Phase Ib/II trials designed to accelerate the development of combinations in several cancer indications by identifying early safety and efficacy signals and establishing clinical data. ¹¹ , ¹² , ¹³ Studies in the MORPHEUS platform include MORPHEUS‐pancreatic cancer, which evaluates immunotherapy‐based combinations in patients with metastatic pancreatic ductal adenocarcinoma (PDAC). Primary results from the MORPHEUS PDAC study evaluating the anti–PD‐L1 antibody atezolizumab (Atezo; F. Hoffmann‐La Roche, Ltd.) in combination with PEGylated recombinant human hyaluronidase (PEGPH20; Halozyme, Inc.), an anti–stromal and extracellular matrix modulator, in patients with pancreatic adenocarcinoma (PDAC) have been previously reported, and demonstrate that the MORPHEUS platform can evaluate new treatment combinations in an efficient and expedited manner. ¹⁴

Data from MORPHEUS PDAC were also used in this ad hoc analysis. Considering the small number of patients enrolled to meet the platform objectives of accelerated development and testing with multiple treatment arms, the trial platform was ideal for this assessment. The Atezo + PEGPH20 RCT experimental arm was compared with an HC arm consisting of patients treated with either folinic acid (leucovorin)‐fluorouracil‐oxaliplatin (mFOLFOX6) or gemcitabine/nab‐paclitaxel (gem/nabP) from the RCT IC arm, supplemented with comparable EC data from a de‐identified EHR‐derived RWD. This ad hoc analysis aimed to test the feasibility of HC implementation and inform how external RWD data could potentially be leveraged to improve treatment‐effect estimation and precision to aid trial decision‐making. The success criteria for this analysis were defined as replication of trial hazards ratio and shorter confidence interval.

METHODS

Data sources

In this analysis, two data sources were used.

Trial data

Trial data was obtained from the Phase Ib/II, open‐label, multicenter, randomized MORPHEUS PDAC platform trial (ClinicalTrials.gov; NCT03193190), which was designed to assess the efficacy and safety of immunotherapy‐based treatment combinations. The trial design (Figure S1 ) and study results were previously reported. ¹⁴ In brief, the study enrolled 169 second‐line patients with metastatic PDAC, of whom 108 patients (N = 66 in the Atezo + PEGPH20 arm, and N = 42 in the IC chemotherapy control arm) received at least one dose of study treatment (Figure S2 ). Patients in the IC arm received either modified FOLFOX6 (mFOLFOX6: 5‐fluorouracil (5‐FU), oxaliplatin, leucovorin) or gemcitabine and nab‐paclitaxel (gem/nabP), depending on what regimen they had received in the 1L setting. The study arm did not meet its primary endpoint for improvement by investigator‐assessed objective response rates per RECIST 1.1, or its secondary endpoints for progression‐free survival (PFS) or overall survival (OS) in the efficacy evaluable population.

Real‐world data (RWD)

The external data was obtained from the Flatiron Health database (Figure 1 ). Both trial and external data included patients with PDAC who progressed after the first line (1L) of 5‐FU or gem‐based chemotherapy in the metastatic setting. Data were collected from patients randomized to the Atezo + PEGPH20 arm or the IC arm. The EC arm was derived from the nationwide Flatiron Health de‐identified EHR‐derived database, a longitudinal database that includes de‐identified patient‐level information from structured data (e.g., laboratory values, prescribed treatments) and unstructured data (e.g., biomarker reports) collected via technology‐enabled chart abstraction from physicians' notes and other documents originated from approximately 280 US cancer clinics (approximately 800 sites of care), primarily community oncology settings. ¹⁵ , ¹⁶ Provisions are established to prevent reidentification to protect patients' confidentiality. Data used in this study covered the period from January 2011 to August 2019. While screening the most appropriate RWD options for use as EC in this study, RWD sources were evaluated in terms of accuracy, completeness of key data elements, availability, and timeliness of key data elements (exposure, outcomes, covariates) and sufficient numbers of representative patients in the EC arm. ¹⁷ , ¹⁸ Raw EHR data collection is based on data collection for clinical documentation, practice management, and billing, and is not standardized in frequency, content, or format across clinical sites, which challenges its use for research. Therefore, strict quality control procedures including consistent computational curation of structured and unstructured data, continuous data abstraction monitoring, extensive abstractor training, benchmarking, and cross‐data comparison are applied by Flatiron Health to create an analyzable data set with maximized data integrity for research use. ¹⁹ , ²⁰ All the key variables required to apply the trial selection criteria were available in the data. If missing data were identified, they were grouped and presented as "other". The sensitivity of the composite mortality endpoint, compared to the national death index, ranged from 85 to 90% across multiple indications. ²¹ Distribution of patient characteristics and level of missingness for patients with PDAC was comparable to the two authoritative sources for population cancer surveillance and research in the US: the Surveillance, Epidemiology, and End Results (SEER) and the National Program of Cancer Registries (NPCR). ¹⁵

Figure 1.

Hybrid control study design. ^aNCT03193190; ^bSeparate models were used for 1:1 PS matching and PS adjustment. Atezo, atezolizumab; EC, external control; gem/nabP, gemcitabine/nab‐paclitaxel; HC, hybrid control; IC, internal control; mFOLFOX6, folinic acid (leucovorin)‐fluorouracil‐oxaliplatin; OS, overall survival; PDAC, pancreatic ductal adenocarcinoma; PEGPH20, pegylated recombinant human hyaluronidase; PS, propensity score; R, randomized.

Construction of the external control arm

The EC arm was built to mimic the trial patients (Figure 2 ). First, to create an EC arm of real‐world patients as comparable as possible to the trial patients, trial inclusion/exclusion criteria were adapted and applied to the external data for which the corresponding variables were available (Table S1 ). These criteria included histologically or cytologically confirmed diagnosis of metastatic PDAC and receiving subsequent treatment after 1L 5‐FU or gem‐based chemotherapy in the metastatic setting (with the first date of treatment with either 5‐FU or gem‐based chemotherapy in the second line as the index date), ≥18 years of age, Eastern Cooperative Oncology Group (ECOG) performance status scores of 0–1 within −30/+7 days of the start of 2L treatment and normal baseline laboratory values within 30 days prior to start of the 2L treatment. Patients with missing ECOG or laboratory test values (absolute neutrophil count, absolute lymphocyte count, albumin, platelet, hemoglobin, alanine aminotransferase, aspartate aminotransferase, alkaline phosphatase, bilirubin, creatine, and calcium) was observed in this dataset (Table S2 ). While we found no impact of missing laboratory values or ECOG on OS (Figure S3 ), we adopted a conservative approach and excluded these patients from the analysis. In addition to these criteria, patients in the EC arm were required to have started their 1L therapy within −14/+90 days of their metastatic PDAC diagnosis to minimize the risk of including patients who may have received treatment outside the real‐world network settings prior to starting treatment within the network, thus avoiding misclassification of the line of therapy. Similarly, patients previously enrolled in clinical trials and those who had received any PDAC systemic treatment within 2 weeks prior to the start of 2L treatment or who had the diagnosis of any secondary primary cancers were excluded. Patients referred to hospice care prior to the start of 2L treatment were omitted to prevent the inclusion of those with limited life expectancy, as they would have been ineligible for the trial.

Figure 2.

Schematic for the inclusion or exclusion of patients into the external control arm from the real‐world database. 1L, first line; 2L, second line; 5‐FU, 5‐fluorouracil; ECOG, Eastern Cooperative Oncology Group; gem/nabP, gemcitabine/nab‐paclitaxel; mFOLFOX6, folinic acid (leucovorin)‐fluorouracil‐oxaliplatin; PDAC, pancreatic ductal adenocarcinoma.

Propensity score (PS) matching

In this study, propensity score (PS) matching was used to balance covariates in the two treatment arms, Atezo + PEGPH20 arm (n = 66) and the EC arm (n = 120). ²² The PS was the conditional probability of receiving Atezo + PEGPH20 (Yes/No, a binary dependent variable). Clinical covariates including age, sex, ECOG performance status, race, history of smoking, stage at first diagnosis, and time from metastatic diagnosis to the start of 2L treatment were added into a multivariable logistic regression model to predict the probability of use of Atezo + PEGPH20. The predicted probability derived from the logistic equation was used as the PS for each individual. The patients in the two groups were then matched by PS, using the nearest matching method in the R package MatchIt with default settings with a 1:1 ratio. The selected patients from the EC arm constituted the matched EC arm and were combined with the IC arm to form the HC arm for further analysis.

Overall survival (OS) analyses

OS was the primary outcome for this analysis, defined as the time from randomization (trial data) or initiation of 2L treatment (external data) to death from any cause or last known date alive. Because the longest follow‐up time of any patients in the trial data was approximately 24 months, patients in the external data who were still alive after 24‐month follow‐ups were censored at 24 months for consistency.

The same OS analysis was performed using all three models described below to estimate the HR comparing Atezo + PEGPH20 arm vs. the hybrid control arm consisting of the IC arm and the PS‐matched EC arm; the arm allocation variable was the sole covariate in the analysis.

Bayesian dynamic borrowing

Bayesian dynamic borrowing approach was used to assess the treatment effect. A 95% credible interval (written as CI as well for ease) is reported. Bayesian dynamic borrowing is a Markov chain Monte Carlo (MCMC)‐driven Bayesian hierarchical survival model with a commensurate prior to control the degree of borrowing information from the external data, which has demonstrated benefits in a previous Phase III HC analysis. ²³ Assuming the survival time $f\left(t\right)$ follows a Weibull distribution, the density function of the jth patient is specified as:

$$f_j\left(t\right)=r{e}^{Y_j\theta }{t}^{r-1}\exp \left(-e^{Y_j\theta }{t}^r\right)$$

where θ is a coefficient vector having one element for each arm (θ _{Atezo+PEGPH20}, θ _IC and θ _EC) and $Y_j$ is the corresponding covariate vector containing indicator variables for each arm for the jth patient. An exponential distribution r ~ Exp(1) was used. All coefficients in θ except the θ_IC were assumed to have independent vague normal priors N(0, 10,000). A commensurate prior ²⁴ was used on θ _IC, assuming θ _IC|θ _EC ∼ N(θ _EC, ɛ ²), where 1/ɛ is a crucial precision parameter with a half‐Cauchy ²⁵ hyper‐prior that controls the degree of borrowing. This JAGS implementation used three parallel MCMC chains, each run for a 1,000‐interaction adaptation and a 5,000‐iteration burn‐in period followed by a 20,000‐iteration production run.

Due to the lack of a gold standard method for HC analysis, alternate OS models were used for sensitivity analysis to interpret the primary results and generate a totality of evidence. ²⁵ , ²⁶ These included the Cox pooled‐control model and frailty Cox model.

Cox pooled‐control model

In this model, patients from the external data and patients from the IC arm were pooled to form a new control arm called the hybrid control. A Cox proportional hazards model was developed to estimate the treatment effect of Atezo + PEGPH20 compared with the pooled control, which is referred to as the Cox pooled‐control model. The hazard $h_j\left(t\right)$ of the jth patient is specified as

$$h_j\left(t\right)=h_0\left(t\right)\exp \left(s{u}_j\right)$$

where $h_0\left(t\right)$ is an unspecified baseline hazard function, $u_j$ indicates if the jth patient was from the Atezo + PEGPH20 arm or the pooled‐control arm.

Frailty Cox model

A frailty model is a Cox model with random effects to account for individual or cohort heterogeneity. If we use cohort i to represent if a patient was from the clinical trial arms (the Atezo + PEGPH20 or the internal control arm) or the external control arm, then the hazard ${\lambda }_{\mathit{ij}}\left(t\right)$ of the jth patient from cohort i is specified as:

$${\lambda }_{\mathit{ij}}\left(t\right)={\lambda }_0\left(t\right)\exp \left(\alpha d_j+b_i\right)$$

where ${\lambda }_0\left(t\right)$ is an unspecified baseline hazard function, $d_j$ indicates if the jth patient was from the Atezo + PEGPH20 arm or a control arm (no matter an internal or external control), and $b_i$~ N(0,${\sigma }^2$) is the shared frailty for each cohort, $i=1,2$.

RESULTS

Hybrid control design

Primary results of patients enrolled in the Atezo + PEGPH20 arm (n = 66) and the IC arm receiving mFOLFOX6 or gem/nabP (n = 42) from the MORPHEUS PDAC study were previously reported ¹⁴ and included in this analysis (Figure 1 ; Figure S2 ). Key inclusion/exclusion criteria from MORPHEUS PDAC were applied to patients from the real‐world cohort, and patients meeting the criteria were included in the study as the EC arm (n = 120) (Figure 2 ).

Balancing baseline confounding variables

Baseline variables were generally balanced among the Atezo + PEGPH20 (n = 66), IC (n = 42), and EC (n = 120) arms; however, the patients in the EC arm were older and had a higher proportion with a history of smoking and shorter time from metastatic diagnosis to the start of 2L treatment (Table 1 ). The distributions of PSs between the trial arms (Atezo + PEGPH20 arm (n = 66) + IC arm (n = 42) = 108) and the EC arm (n = 120) also presented a considerable overlap (Figure S4 ), supporting the similarity between the two data sources.

Table 1.

Baseline demographics of MORPHEUS PDAC vs. real‐world cohort

	MORPHEUS PDAC		Real‐world cohort
	Experimental arm	IC arm	EC arm	Matched EC arm
	Atezo + PEGPH20, n = 66	mFOLFOX6 or gem/nabP, n = 42	mFOLFOX6 or gem/nabP, n = 120	mFOLFOX6 or gem/nabP, n = 66
Age, median (IQR), years	60.00 (51.50, 66.00)	61.50 (54.25, 69.00)	65.00 (59.00, 71.25)	61.50 (57.00, 67.00)
Time from metastatic diagnosis to start of 2L treatment, median (IQR), months	9.63 (4.06, 14.32)	8.42 (5.74, 11.89)	6.92 (3.71, 10.90)	7.58 (3.94, 12.45)
Sex, n (%)
Female	27 (40.9)	20 (47.6)	57 (47.5)	25 (37.9)
Male	39 (59.1)	22 (52.4)	63 (52.5)	41 (62.1)
ECOG performance status, n (%)
0	25 (37.9)	17 (40.5)	43 (35.8)	25 (37.9)
1	41 (62.1)	25 (59.5)	77 (64.2)	41 (62.1)
Race, n (%)
White	41 (62.1)	27 (64.3)	87 (72.5)	41 (62.1)
Othera	25 (37.9)	15 (35.7)	33 (27.5)	25 (37.9)
Documentation of smoking history, n (%)
Yes	27 (40.9)	19 (45.2)	64 (53.3)	29 (43.9)
No	39 (59.1)	23 (54.8)	56 (46.7)	37 (56.1)
Stage at the first diagnosis, n (%)
IV	51 (77.3)	31 (73.8)	86 (71.7)	48 (72.7)
Othera	15 (22.7)	11 (26.2)	34 (28.3)	18 (27.3)
Index year, n (%)
2011–2013	0	0	0	0
2014–2016	0	0	41 (34.2)	22 (33.3)
2017–2019	66 (100)	42 (100)	79 (65.8)	44 (66.7)

1L, first line; 2L, second line; Atezo, atezolizumab; EC, external control; ECOG, Eastern Cooperative Oncology Group; gem/nabP, gemcitabine/nab‐paclitaxel; IC, internal control; IQR, interquartile range; mFOLFOX6, folinic acid (leucovorin)‐fluorouracil‐oxaliplatin; PDAC, pancreatic ductal adenocarcinoma; PEGPH20, PEGylated recombinant human hyaluronidase.

^a Other includes unknown/missing.

PS matching between the Atezo + PEGPH20 arm (n = 66) and EC arm (n = 120) was performed to further form a matched EC arm (n = 66). The distribution plot showed that almost every patient in the Atezo + PEGPH20 arm found a matched patient in the EC arm with a small PS difference (Figure S5A ). The balance of baseline variables was also improved after the PS matching (Table 1 , Atezo + PEGPH20 arm vs. matched EC arm; Figure S5B ).

Assessment of OS using the HC arm

The hazard ratios (HRs) of the Atezo+PEGPH20 arm (HR 0.91, 95% CI: 0.56, 1.49), EC arm (HR 0.80, 95% CI: 0.51, 1.25), and matched EC arm (HR 0.81, 95% CI: 0.50, 1.32) compared with the IC arm were nonsignificant. Moreover, the unadjusted median OS (mOS) across the IC arm (6.80 months), Atezo + PEGPH20 arm (7.06 months), EC arm (7.13 months), and matched EC arm (7.13 months) were similar (Figure 3 ).

Figure 3.

Overall survival in the hybrid control study. Atezo, atezolizumab; EC, external control; IC, internal control; mOS, median overall survival; PEGPH20, pegylated recombinant human hyaluronidase.

The direction and magnitude of HRs from the multiple HC analyses were comparable with the trial HR excluding the external data (Figure 4 ). Compared with the formal trial results (HR 0.91; 95% CI: 0.56, 1.49), numerical differences in HR estimates were observed, while the 95% CI was shortened after incorporating the external data into analyses. Among the three OS models using the matched EC arm, the Cox pooled‐control model and the frailty Cox model provided identical results (HR 1.03; 95% CI: 0.70, 1.51), while the Bayesian dynamic borrowing results presented a similar HR estimate with a smaller 95% CI decrease (HR 1.02; 95% CI: 0.63, 1.51).

Figure 4.

Direction and magnitude of HRs from the multiple HC analyses vs. the trial HR without external data. Atezo, atezolizumab; CI, confidence interval for Cox pooled‐control model and frailty Cox model while credible interval for dynamic borrowing model; EC, external control; HC, hybrid control; HR, hazard ratio; IC, internal control; PEGPH20, pegylated recombinant human hyaluronidase; PS, propensity score.

DISCUSSION

The MORPHEUS PDAC trial evaluated immunotherapy‐based combinations, including the efficacy and safety of Atezo + PEGPH20 compared with mFOLFOX6 or gem/nabP. One of the limitations of this small Phase Ib trial was the limited sample size, especially for the time‐to‐event endpoint, OS. This limitation can be mitigated by utilizing the HC arm, which was one of the motivations for conducting the ad hoc analysis to supplement the trial.

Important aspects of using ECs and designing HC arms is the minimization of biases that could affect the interpretation of the results, and the availability of high‐quality RWD that can mirror the trial population as closely as possible. Burger et al described several key sources of bias when using external controls. ²⁷ Endpoint and assessment biases were limited as we objectively assessed OS. To minimize selection bias, we applied five out of nine trial inclusion criteria, specific to disease and treatment characteristics, ECOG, and laboratory values, which were identified to have the largest impact of hazard ratio in OS analysis. While the Flatiron Health database is rigorously checked for data quality, ¹⁹ , ²⁰ we encountered missing data, primarily with ECOG (32.1%) and laboratory values (40.1%; See Table S1 ). For laboratory values, missingness seemed to be dependent on temporal variability in data quality instead of patient characteristics. Yet, there were no differences in key measured baseline characteristics or OS with known vs. missing laboratory values (Figure S3 and S4 ), indicating that there was no impact on the missing laboratory values. This was also true for ECOG scores (data not shown). Despite this, we adopted a conservative approach to exclude these patients from the analysis. The rest of the inclusion/exclusion criteria could not be applied due to limited reporting of comorbidities, non‐oncology treatment, and clinical characteristics in routine clinical practice (Table S2 ). After applying these criteria, only a limited portion of patients had missing race or stage at first diagnosis; those patients' missing data were grouped into the category “other” and did not have a big impact on analyses. Regarding missing RCT data, missing values of the stage at the first diagnosis for seven patients (four in the Atezo + PEGPH20 arm and three in the IC arm) were coded as “other.” Seven patients in the Atezo + PEGPH20 arm had missing values in time from metastatic diagnosis to the start of 2L treatment.

Usually, contemporaneous controls are preferable to historical controls to mitigate calendar time biases. However, the historical control (2011–2019) instead of a contemporaneous control (2017–2019, trial duration) was used in this study for the following reasons: (a) the global standard of care for 2L PDAC treatments inclusive has not changed inclusive of the real‐world dataset period (2011–2019) used for this analysis, and (b) restricting the RW index period to 2017–2019 would yield only 77 patients as external control, which may limit the potential of analyses. Given these considerations, we selected the historical control strategy. ²⁸ Admittedly, the Flatiron Health database is based on a US population only, whereas MORPHEUS PDAC is a global study that enrolled only 56% of 2L internal control and 59% of the Atezo + PEGPH20 patients from the US. The MORPHEUS trial was also an open‐labeled study with frequent assessments indicated for safety evaluations; this likely provided a level of care and compliance from both the patients and investigators that differed from the standard community practices that would be reflected in the Flatiron database. Therefore, there may be some regional and study biases that may impact OS that are not accounted for in this analysis.

Although we have performed the PS matching to create the matched EC arm, there was still a limited difference between the IC and matched EC arms (Table 1 , Figure S5B and the HR of the matched EC vs. IC arm (HR 0.81, 95% CI: 0.50, 1.32)), which may lead to a bias in the OS analyses. To address this concern, we applied multiple analytical approaches when utilizing the matched EC arm. A Bayesian dynamic borrowing method was included because it can adaptively downweight EC to a desired level by adopting sensible hyper‐priors. The sensitivity analysis included the frailty Cox model to approach possible heterogeneity between the trial data and the external data. The frailty Cox model showed nearly identical results to the Cox pooled‐control model, which may result from the similarity between the IC and EC arms. One possible limitation of the frailty model in this work is that the variance parameter may be underestimated because the frailty model has only two levels. ²⁹ The HR point estimates of the Bayesian dynamic borrowing model were closer compared to all other approaches to the HR estimates without any EC data. However, the CIs from the Bayesian dynamic borrowing model were also wider compared with the frailty Cox model. This may be explained by the prespecified conservative hyper‐prior choice, which led to less information borrowing, for this Bayesian framework. Sensitivity analyses using PS as a covariate were also conducted and yielded similar results to those PS 1:1 matching (results not included).

The analysis methods and operating procedures outlined in this study may be generalized for future projects of a similar nature. Such trials can be smaller, and/or unequal randomization may be used to place proportionately more patients in the experimental treatment arm of a study, potentially increasing the amount of information obtained on the efficacy and safety of novel treatments. This approach may also make trials more attractive to patients and decrease the likelihood of dropout. While this analysis is a good start, there are a lot of considerations when trying to justify the use of HC arms to draw conclusions about treatment efficacy.

The results of this study should be interpreted with several limitations. The RWD set draws exclusively from the US patient population, which limits patient diversity. Additionally, RWD‐derived EC arms cannot act as a control for unmeasured or unknown confounders and are restricted to known confounders that are available in both the trial data and the EHR database. Not all inclusion/exclusion criteria from the trial could be applied to the RWD‐derived patients due to the unavailability of some of the variables in the RWD (e.g., life expectancy, criteria based on secondary metastasis). It is noteworthy that despite this limitation, the EC arm replicated the OS output from the IC arm in this analysis. Another limitation was that other endpoints such as objective response rate (ORR), PFS, and adverse events could not be assessed due to the limited ability to validate this information in the RWD.

This analysis demonstrated the value of leveraging RWD in modern clinical trial design to supplement trial control arms. The study offered an actual application of an HC arm from an ongoing clinical trial and was not based on hypothetical assumptions or simulations. Future research expanding this foundational work to other patient populations, treatment settings, indications, and endpoints (ORR, PFS, toxicity) could help increase the utility of trials with an HC arm to inform decision‐making in early development programs such as MORPHEUS. Numerous regulatory and reimbursement agencies have indicated an openness to exploring the use of completed RCTs or RWD as support for sparse control‐arm data for comparative efficacy research. ³⁰ , ³¹ Cumulative evidence from similar analyses can help advance these discussions from the conceptual to the applied.

FUNDING

This manuscript was sponsored by F. Hoffmann‐La Roche Ltd. and Genentech, Inc.

CONFLICT OF INTEREST

B.C., S.L., M.T.B., I.R.‐R., J.Z., and J.L. are employees and shareholders of Genentech, Inc./F. Hoffmann‐La Roche Ltd. S.K.M., X.Z., and J.P. were employees of Genentech, Inc. R.M. was an employee and shareholder of Roche Products Ltd. M.A.P. received funding from a research agreement between Genentech, Inc./F. Hoffmann‐La Roche Ltd and the University of North Carolina. A.H.K. received research support for the conduct of this clinical trial paid directly to his institution. A.H.K. receives grants from Abgenomics, Celgene, Pancreatic Cancer Action Network, Parker Institute for Cancer Immunotherapy, and Verastem paid directly to his institution. A.H.K. received a payment or honoraria from Fibrogen. A.H.K. participates in data safety monitoring boards for Genentech, Grail, and Ipsen. A.H.K. serves as an advisory board member for Arcus, Fibrogen, and Merus. A.H.K. has a leadership or fiduciary role at the National Cancer Institute. D.‐Y.O. receives research grants from Array, AstraZeneca, BeiGene, Eli Lilly, Handok, MSD, and Novartis. D.‐Y.O. serves as a consultant/advisory board member for Abbvie, Arcus Biosciences, ASLAN, AstraZeneca, Astellas, Basilea, Bayer, BeiGene, BMS/Celgene, Eutilex, Genentech/Roche, Halozyme, Idience, IQVIA, LG Chem, J‐Pharma, Merck Serono, Mirati Therapeutics, Moderna, MSD, Novartis, Taiho, Turning Point, Yuhan, and Zymeworks. M.P.‐S. received research support for the conduct of this clinical trial paid directly to his institution. M.P.‐S. receives research grants from AstraZeneca and Novocure. M.P.‐S. serves as a consultant/advisory board member for AstraZeneca and Taiho. M.P.‐S. receives travel support from AstraZeneca, BMS, and Oncosyl. M.P.‐S. participated in a data safety monitoring board or advisory board for AstraZeneca.

AUTHOR CONTRIBUTIONS

A.H.K., D.‐Y.O., S.K.M., B.C., R.M., S.L., M.T.B., I.R.‐R., J.Z., X.Z., J.L., J.P., M.A.P., and M.P.‐S. wrote the manuscript. S.K.M., D.‐Y.O., B.C., J.Z., R.M., M.T.B., X.Z., J.L., J.P., S.L., I.R.‐R., and M.A.P. designed the research. A.H.K., D.‐Y.O., and M.P.‐S. performed the research. B.C., D.‐Y.O., I.R.R., and R.M. analyzed the data.

Supporting information

Table S1.

Table S2.

Figure S1.

Figure S2.

Figure S3.

Figure S4.

Figure S5.

DATA AVAILABILITY STATEMENT

Qualified researchers may request access to individual patient‐level data through the clinical study data request platform (https://vivli.org/). Further details on Roche's criteria for eligible studies are available here (https://vivli.org/members/ourmembers/). For further details on Roche's Global Policy on the Sharing of Clinical Information and how to request access to related clinical study documents, see here (https://www.roche.com/research_and_development/who_we_are_how_we_work/clinical_trials/our_commitment_to_data_sharing.htm). The data that support the findings of this study were originated by Flatiron Health, Inc. Requests for data sharing by license or by permission for the specific purpose of replicating results in this manuscript can be submitted to publicationsdataaccess@flatiron.com to determine licensing terms.