Clinical trials give patients hope — especially when all other available treatments have been exhausted. That hope compels patients to commit a substantial amount of time and effort for the chance to receive a potentially effective treatment before it goes to market. In turn, clinical research needs patients. Without adequate patient enrollment, there isn’t enough data for researchers to evaluate the safety and efficacy of novel therapeutics, nor for regulators to approve them.
Recruitment failure is one of the main reasons clinical trials are canceled or delayed.1 Nearly 80% of clinical trials fail to meet their enrollment timelines, and up to 50% of research sites enroll one or no patients.2 In a study of patients who declined clinical participation, 36% cited issues with the study protocol issues, including a distaste for placebo-controlled studies.3 This finding highlights a need to run trials designed to assuage fears about placebo arm assignment.
Methods that use existing patient data from previously completed clinical trials are desirable not only for reducing trial sizes but also for honoring the valuable contributions of past trial participants. Both prognostic digital twins* and external control arms (ECAs) incorporate external data into clinical trials to reduce trial sizes, but only digital twins overcome the issues of confounding variables and bias.4,5
Digital twins: smaller trials that generate regulatory-acceptable evidence
Digital twins are comprehensive, longitudinal predictions of the clinical trajectory of a patient under the control condition. They are used in randomized controlled trials called, TwinRCTsTM, to enable smaller control arms by up to 35% without introducing bias.6
The first step in a TwinRCT involves training and validating a machine learning model on historical patient data.7 Next, baseline data collected during a patient’s first site visit is fed into this pre-trained model to generate that patient’s digital twin (see figure below), and this process is repeated for every patient in the trial prior to their randomization to the active treatment (experimental drug) or control (placebo) arm of the trial. Lastly, prognostic scores derived from digital twins are incorporated into the primary trial analysis using Unlearn’s regulatory-qualified Prognostic Covariate Adjustment (PROCOVA™) framework.8
After training the machine learning model, it is ready to create a digital twin using baseline data measured from a specific patient in the clinical trial.
Digital twins and PROCOVA leverage all of the long-established positive characteristics of randomization (removal of bias) and ANCOVA (variance reduction), making them uniquely powerful tools for increasing efficiency in clinical trials across all stages of development. TwinRCTs with digital twins significantly shorten trial timelines, helping to bring novel treatments to patients sooner.
ECAs introducing confounding variables and bias
ECAs are populated with patients who were not in the original patient sample of the trial. Instead, the control group is populated with patients who are selected from an external data source so that it appears as if they were randomized into the study along with the other patients. Each patient in the external control population is assigned a propensity score computed from a patient’s baseline characteristics, describing the probability of that patient belonging to the active treatment group given the covariates. Then, a subset of patients from the external control population with similar propensity scores to those in the study population is selected to create an ECA.
The most important assumption underlying propensity score matching for creating ECAs is that there are no unmeasured confounding variables. Confounders are unmeasured factors that are not included in the study but can still influence the study’s outcome, making it difficult to draw conclusions about the drug’s efficacy. Some examples include alcohol and cigarette usage, diet, medications, and pre-existing conditions that might interfere with the study’s disease indication. Susceptibility to confounding increases the probability of producing a false positive result (specifically, the Type-I error rate). Since ECAs can’t completely control for bias, their limitations make them inappropriate for many studies.
Digital twins move research forward
ECAs and digital twins are two methods for incorporating external data into clinical trials, enabling control arms that require fewer patients. However, a key difference between them is that external data used to train digital twin-generating models never actually enters the trial. This difference influences the statistical guarantees of each procedure. TwinRCTs control for bias and produce regulatory-acceptable evidence across all disease indications. ECAs, on the other hand, introduce bias and therefore have achieved regulatory approval only on a case-by-case basis.
Our industry has a responsibility to nurture hope in clinical research— the future of medicine depends on it. By running trials that give patients a higher probability of receiving the experimental treatment, providers can help eliminate one more barrier between patients and better health outcomes. Unlike ECAs, only TwinRCTs with digital twins help to overcome participation challenges while generating the evidence needed to move research forward.
Read Unlearn’s whitepaper about how digital twins provide better statistical outcomes compared to ECAs in clinical trials.
*Prognostic digital twins are predictions of control outcomes for each individual patient enrolled in an RCT. For the duration of this article, we use the shortened term “digital twins” to mean “prognostic digital twin”.