Next Generation Clinical Trial Design: Applying Cohort Optimization & Synthetic Control Arms
Clinical trials drive much of the significant cost and time needed to bring a drug to market. A study published in JAMA Internal Medicine in late 2018 showed that the median cost per patient for a clinical trial is over $41,000. Additionally, a study in Nature Reviews Drug Discovery from 2019 showed that over 93% of drug candidates entering Phase I fail to make it to market.
These high failure rates and increasing per-patient costs make conducting pivotal studies a daunting prospect for even the most deep-pocketed pharmaceutical companies. The current COVID-19 crisis has added a formidable exposure risk to the challenges of conventional patient recruitment and trial execution, forcing many companies to pause or even cancel existing trials.
As a result, industry innovators are pioneering new ways to design and execute clinical trials. Recent advances in data analytics applied to medical Real World Data (RWD) have yielded new statistical methods such as Cohort Optimization and Synthetic Control Arms that dramatically reduce the time, cost, and risks associated with clinical trials. Regulators have also shown willingness to embrace some of these new approaches, with the FDA recently soliciting industry feedback on best practices for using machine learning and artificial intelligence (ML/AI) and advanced analytics to drive better clinical outcomes. Companies investing in this innovation gain a significant long-term competitive advantage.
Cohort Optimization
Conventional clinical trial designs typically lean heavily on prior trials to mimic their structure. While modeling trial design around published, well-respected randomized clinical trials (RCT) has clear advantages, the resulting designs will likely experience similar biases, inefficiencies, and limitations. Cohort Optimization takes advantage of risk models that accurately predict clinical events of interest. These models are derived from individual subject data in published RCTs or from the application of ML/AI techniques to RWD. The models are used to define patient groups which are more likely to experience events / outcomes of clinical interest.
The knowledge of which patient groups are more likely to experience clinically relevant outcomes is used to define inclusion / exclusion criteria which can enrich clinical trial arms for rate of events of interest. The resulting trials can achieve the same level of statistical significance with smaller trial arms and a shorter trial period. This method of design helps reduce the risk that a trial will miss the number of events needed to prove a treatment effect, and will improve a study’s ability to target composite outcomes by modeling these outcomes in the initial design phase. A longitudinal real world dataset provides a means of performing a “dry run” study by using historical data to gauge event / outcome rates and perform further patient cohort profiling.
The value of these methods is getting broader recognition in the industry and regulators. Just last year, the FDA issued guidance on the use of enrichment strategies in clinical studies. This class of methods has, in fact, been used in a number of disease areas. One example was published in the Alzheimer’s Association journal Alzheimer’s & Dementia: Translational Research & Clinical Innovations by H. Lundbeck A/S and several academic groups. They showed that, to achieve >80% statistical power to detect a 2-point difference on a standard cognitive test at 0.05 level of significance, cohort optimization methods with Alzheimer’s biomarkers could reduce the number of patients needed per arm by up to a factor of 10x!
Reducing the number of patients needed to see a treatment effect has significant implications for the drug industry on cost, risk, and ability to execute a successful clinical trial. Especially for complex disease states or therapeutic areas which have traditionally needed substantial trial recruitment, such an approach will bring novel therapeutics to market faster.
Synthetic Control Arms
If RWD can be used to identify individuals more likely to experience an event based on historical patients like them, then RWD also holds value as a living dataset that can track the impact of a specific therapy over time. Synthetic Control Arms apply statistical methods to RWD to identify a patient population matched to a recruited experimental arm. This matched population functions as a control arm, providing statistical rigor that can match that of a conventional RCT, but selected and monitored using RWD without ever being formally recruited for the study.
This enables trials to recruit substantially fewer patients, potentially avoiding the need to recruit conventional control arms entirely, lessening trial costs and recruitment time. Because synthetic control arms are oftentimes drawn from datasets far larger than what a single study site / contract research organization can reach, they can be constructed to potentially be more representative of a broader population or more tightly matched with an experimental arm than can be achieved by conventional recruiting and randomization.
Synthetic control arms have already been successfully used by forward-thinking companies commercially. One well known example is Roche’s use of a synthetic control arm in a study for its lung cancer drug Alecensa (alectinib). The data for the synthetic control arm was derived from a Real World Dataset powered by an oncology medical record. By leveraging a synthetic control arm (see figure from paper below), Roche was able to accelerate a coverage decision in 20 EU countries by 18 months relative to relying on a conventional Phase 3 study to complete.
BioMarin has another example of a successful use of a synthetic control arm in a clinical trial for its drug Brineura (cerliponase alfa) for pediatric patients with a form of Batten’s disease, a rare inherited neurological disease. In an open label study published in the New England Journal of Medicine in 2018, BioMarin compared outcomes between a conventional experimental arm given Brineura and a synthetic control derived from lookback data in the Dementia in Children (DEM-CHILD) database. The drug received approval from the FDA in April 2017 on the strength of the comparison and shows the applicability of synthetic controls to a wider range of therapeutic areas as well as its utility in settings where trial recruitment is especially difficult (as it is with rare inherited diseases).
Keys to Success
Cohort Optimization and Synthetic Control Arms can, separately or in concert, serve as cost-and-time saving pillars for a next-generation clinical trial strategy. However, to successfully deploy these methods, companies running clinical trials will need to be certain that they:
- Leverage a large real-world dataset: Both Synthetic Control Arms and Cohort Optimization are best carried out using large real-world datasets with significant longitudinal data and sufficient clinical features for meaningful outcome tracking and inclusion/exclusion criteria filtering. This is critical, as more patient data will provide clinical trial designers with a broader population with which to perform a Synthetic Control Arm match as well as to study how different cohort definitions impact incidence of events of interest. Smaller datasets may work if they are focused on a specific disease area, but will struggle in use cases needing larger arms or broader clinical modeling.
- Possess deep analytics capability: To successfully carry out Cohort Optimization and Synthetic Control Arms, clinical trial designers and data teams will need significant analytical capability. The statistical weighting, modeling, arm matching, and lookback analyses are significant undertakings requiring prior expertise. Additionally, the ability to parse large real-world datasets necessitates handling issues around data quality, and mapping available data to specific events and cohort definition criteria is extremely challenging technically and analytically.
- Incorporate RWD-based methods in trial planning: Cohort Optimization and Synthetic Control Arms will have a more transformative impact on clinical trial cost and operation if they are thoughtfully incorporated into trial planning. This necessitates careful consideration of cohort size selection to balance the benefits of cost / time reduction with achieving statistical power and broad population representation. The successful incorporation of these methods in trial planning also requires deep integration with RWD to (1) capture and evaluate events of interest, (2) define and design inclusion/exclusion criteria, and 3) pre-validate hypotheses based on historical cohorts.
Therapeutic companies embracing these new statistical methods can gain a significant competitive edge in the market. To do this, they should seek partners who have access to large-scale RWD, significant experience working with drug makers to analyze and extract insights from RWD, and have rich tooling to facilitate the planning and use of these technologies that can augment existing data science and clinical trial capabilities.
One such partner is HealthPals, a Silicon Valley-based company that developed CLINT™, a precision insights platform capable of understanding clinical RWD from the patient to the medical system. The platform was built with scale in mind, has been run on over a Billion patient-life-years of data, and has been used by major life sciences companies and medical payers to profile user-defined patient cohorts and model their clinical outcomes and disease progression. CLINT™ Cohort Optimization and Synthetic Control Arm offerings support the next-generation of clinical trial design and can work in conjunction to enable significant time and cost savings for its partners. If you are interested in learning more about HealthPals and our CLINT™ platform, contact us at hello@healthpalsinc.com.
