Skip to content

BioCentury – Broadening role for external control arms in clinical trials

BioCentury – Broadening role for external control arms in clinical trials

External control arms are moving from theory to practice as drug developers begin to use them to make internal go/no-go decisions for clinical programs and to support regulatory applications. The field is largely split between those drawing on past clinical trials versus real-world data, and at least one company is pushing the approach further by simulating artificial patients to augment control arms in complex, chronic indications like Alzheimer’s disease.


Randomizing patients into experimental and control arms is critical to building confidence that the benefits observed in patients treated with a drug are due to the drug itself, and not the patients’ baseline characteristics. But the prospect of being assigned to a control group leads many patients to opt out of the clinical trial process all together, and filling control groups in rare and severe diseases is often infeasible or unethical.


External control arms offer a way to reduce the number of study participants treated with placebo or standard of care, decreasing trial size, duration and cost and incentivizing patient participation.


The idea is to replicate traditional randomization using control data from past clinical trials or real-world data (RWD) from electronic health records (EHRs) and other sources. RWD goes beyond natural history data because it can include patients undergoing treatment in real time and provides more detailed information at the level of individual patients.


At a Friends of Cancer Research meeting in November, FDA officials voiced cautious optimism for external control arms, identifying single-arm trials as the setting where the approach has the clearest benefit. The agency has already made a handful of regulatory decisions supported in part by external control data.


Examples include the 2018 approval of PD-L1 inhibitor Bavencio avelumab from Pfizer Inc. and Merck KGaA for Merkel cell carcinoma, where short patient survival times precluded recruitment of a prospective control group. The control arm used data from electronic medical records obtained in community and academic centers.


Another was the April label extension for breast cancer drug Ibrance palbociclib from Pfizer to include men with hormone receptor (HR)-positive, HER2-negative advanced or metastatic breast cancer. The expansion was supported by data from Pfizer’s global safety database, IQVIA claims data and EHRs from the Flatiron Health unit of Roche.


Pfizer Chief Development Officer Chris Boshoff said that for external control arms to go mainstream, they will have to build credibility through repeated use in guiding internal go/no-go decisions, like whether to advance a candidate to Phase II.


While some see any departure from prospective randomized control trials (RCTs) as dangerous, there is growing support for the idea that industry should make as much use of available data as possible to at least augment traditional approaches, particularly in settings that where RCTs are unattainable.


The approach is gaining visibility through company partnerships, pre-competitive workshops and conference presentations, including recent ones by FDA’s Project Switch, a program launched last year to study external controls derived from legacy trial data (see “Clinical Trial and Regulatory Efficiency at ASCO19”).


The most established strategies populate external control arms using previously collected data, taking care to match the baseline characteristics of the patients in the external dataset to those in the new study using a method known as propensity scoring (see Sidebar: “Random by Design”).

Unlearn.AI Inc. is going a step further, using data from historical trials and patient registries to build algorithms that simulate artificial patients. Its simulations have not yet been used to augment control arms in trials.


Unlearn.AI simulates patient trajectories in neurodegenerative and inflammatory diseases based on control group data from dozens of trials. The method learns the probability distributions underlying the historical training data, and uses those distributions to generate new data.


CEO Charles Fisher said Unlearn.AI chose to focus on Alzheimer’s and inflammatory conditions like rheumatoid arthritis and multiple sclerosis because the complexity of the indications is a good match for its machine learning tools, and because it was able to access sufficient historical data through non-profits like Vivli and Transcelerate Biopharma Inc.


Unlearn.AI’s simulations are being used to guide trial design by simulating hypothetic outcomes under different study conditions. In collaboration with an undisclosed pharma, the company has used its model to “ask questions about trial design, size and inclusion criteria” for Alzheimer’s, said Fisher.


The company aims to use its simulations to supplement control arms, and is preparing to seek guidance from regulatory agencies on using its platform to run synthetic controls for Alzheimer’s trials.


Unlearn.AI also hopes to generate “digital twins” of individual patients in experimental arms. “We can ask, what would have happened to this person if they had received the placebo,” said Fisher.


For input data, the company sticks to commonly collected clinical and demographic parameters such as neurological exam scores, APOE mutation status and background medication, but is keeping an eye on newer molecular biomarkers of disease. “We definitely are interested in including those as they become more popular, but we have to see which ones end up becoming adopted,” Fisher said.


The two-year old company has raised a $4.2 million seed round from DCVC Bio, Data Collective, Mubadala Ventures and angel investors. Unlearn.AI also brings in revenue by offering data standardization as a service, and uses that data to further train its models.


External control arms based on control data from previous trials have the advantage of providing high quality, directly measured endpoint and covariate data from extensively monitored global sites, said Ruthanna Davi, VP of data science in Medidata Solutions Inc.’s Acorn AI unit. “These trials are designed to be used for regulatory purposes in the first place.”


These types of external controls are best suited for well-studied diseases where the standard of care has remained constant over time and outcomes are predictable, said Antoine Yver, EVP and Global Head of Oncology R&D at Daiichi Sankyo Co. Ltd.


An example is small cell lung cancer (SCLC), the focus of Project Data Sphere LLC’s external control arms program. Following a 2018 symposium with FDA, the oncology-focused non-profit put out calls to pharmas for SCLC control data; it has received six data sets so far from Amgen Inc., Bayer AG, Eli Lilly & Co. and the Alliance for Clinical Trials in Oncology, and is due to receive another four by the end of the year.


Project Data Sphere is starting to make this external control data available to trial sponsors through its data sharing platform.

“These trials are designed to be used for regulatory purposes in the first place.” – Ruthanna Davi, Medidata Solutions

Program Manager Dave Handelsman said a team led by Dana-Farber Cancer Institute researchers has been standardizing the data, which is due to be published this fall. Project Data Sphere plans to discuss the data with FDA, and Handelsman hopes the conversations will provide greater clarity on how FDA thinks companies should deploy external control arms.


Medidata, which has stored and analyzed data from more than 17,000 clinical trials over the last 20 years, provides drug developers with external control data from previous trials in a range of indications.


The company offers its external control arms as a guide for internal go/no-go decisions, and believes they could eventually play a role in regulatory submissions.


Co-founder and president Glen de Vries told BioCentury the company is investing in methods to standardize data across trials at scale, including machine-learning based approaches.


However, Yver cautioned that submitting past clinical trial data for FDA review as an external control arm to a new study is not as straightforward as it sounds.


FDA requires sponsors to submit patient-level data, but sponsors only have permission to share this past trial data with FDA in aggregate because patients did not consent to having their data used outside of the original study.


“It seems very trivial, but it’s actually very real,” he said, adding that work to de-identify control data and make it available for FDA analyses is ongoing.


In cases where there is limited clinical trial experience with a disease or disease subtype, RWD may be a more fruitful source for external controls.


“So many studies are aimed at rare mutations that may not even have been tested for before the last two years. So all my clinical trial data from more than two years ago is irrelevant,” said Edward Stepanski, SVP and COO of Concerto HealthAI. He said RWD also better captures the baseline characteristics of patients treated in community practices, which can be substantially different from those of patients in clinical trials.


The concern is that variability in data quality and in reporting of endpoints and covariates could confound analyses.


CEO Jeff Elton said Concerto HealthAI focuses on mining unstructured oncology data from EHRs, including physicians’ notes and genetic data, which he said is key to achieving the same rigor as RCTs. “Most datasets that most people use are composed of the structured fields alone.”


The company has undisclosed projects underway using external control arms to support regulatory submission. In March and April, Concerto HealthAI announced deals with Bristol-Myers Squibb Co. and Pfizer Inc., respectively, to conduct external control arm and prospective real-world outcomes studies in cancer.


Flatiron Health has already used RWD-based external controls to support regulatory submissions. Its platform combines curated oncology EHR data with tumor mutation profiling data from Roche’s Foundation Medicine unit.


In addition to supporting FDA’s label extension for Ibrance, standard-of-care data from Flatiron’s platform helped parent company Roche gain speedier European market access from health technology assessment authorities for its cancer drug Alecensa alectinib in ALK-positive advanced non-small cell lung cancer (NSCLC) in patients who failed or were intolerant to crizotinib.


In 2017, Alecensa gained conditional approval in the indication in Europe based on Phase II data from 225 patients treated with the drug and RWD from 77 patients treated with ceritinib.


Roche acquired Flatiron in 2018 for $1.9 billion (see “Accelerating Flatiron”).


Flatiron was not available to comment in time for publication.


monARC Bionetworks is building RWD-based external control arms for rare diseases like idiopathic pulmonary fibrosis by reaching out to patients directly. The company builds disease-specific patient research networks via advocacy groups, health systems and social media.


“We provide them tools to aggregate all of their health records” and track their symptoms, said CEO and founder Komathi Stem. “They also want to participate in research. Either they want to be matched to a potential trial, or they just want to share their data.”


She said monARC’s longitudinal database can incorporate data from any marketed test, and in the future could include inputs from wearables like sleep monitors. The company can also leverage its direct patient interactions to do prospective studies, such as surveys on patient-reported outcomes (PROs), said Stem.


monARC has not announced external control arm partnerships with drug developers.


Yver thinks RWD-based external controls will be important for settings where patients frequently cross over from the control group into a new line of treatment, because patients in real-world settings are less likely to confound the study by switching away from standard-of-care than patients in clinical trials.…-?kwh=broadening%3C%7C%3Erole%3C%7C%3Eexternal%3C%7C%3Econtrol%3C%7C%3Earms%3C%7C%3Eclinical%3C%7C%3Earm%3C%7C%3Econtrols%3C%7C%3Etrials%3C%7C%3Etrial%3C%7C%3Ebroaden%3C%7C%3Eclinic