Real-time adaptive randomization of clinical trials

Objectives

To evaluate real-time (day-to-day) adaptation of randomized controlled trials (RCTs) with delayed endpoints – a “forward-looking optimal-experimentation” form of response-adaptive randomization. To identify the implied tradeoffs between lowered mortality, CIs, statistical power, potential arm misidentification, and endpoint rate change during the trial.

Study Design and Setting

Using data from RCTs in acute myocardial infarction (30,732 patients in the Global Utilization of Streptokinase and Tissue Plasminogen Activator for Occluded Coronary Arteries, GUSTO-1) and coronary heart disease (12,218 patients in the EURopean trial On reduction of cardiac events with Perindopril in stable coronary Artery disease, EUROPA), we resample treatment-arm assignments and expected endpoints to simulate (1) real-time assignment, (2) forward-looking assignments adapted after observing a fixed number of patients (“blocks”), and (3) a variant that balances RCT and real-time assignments. Blinded real-time adaptive randomizations (RTARs) adjust day-to-day arm assignments by optimizing the tradeoff between assigning the (likely) best treatment and learning about endpoint rates for future assignments.

Results

Despite delays in endpoints, real-time assignment quickly learns which arm is superior. In the simulations, by the end of the trials, real-time assignment allocated more patients to the superior arm and fewer patients to the inferior arm(s) resulting in less mortality over the course of the trial. Endpoint rates and odds ratios were well within (resampling) CIs of the RCTs, but with tighter CIs on the superior arm and less-tight CIs on the inferior arm(s) and the odds ratios. The variant and patient-block-based adaptation each provides intermediate levels of benefits and costs. When endpoint rates change within a trial, real-time assignment improves estimation of the end-of-trial superior-arm endpoint rates, but exaggerates differences relative to inferior arms. Unlike most response-adaptive randomizations, real-time assignment automatically adjusts to reduce biases when real changes are larger.

Conclusion

Real-time assignment improves patient outcomes within the trial and narrows the CI for the superior arm. Benefits are balanced with wider CIs on inferior arms and odds ratios. Forward-looking variants provide intermediate benefits and costs. In no simulations, was an inferior arm identified as statistically superior.

Plain Language Summary

Randomized controlled trials (RCT) are the gold standard in clinical trials — typically half of the patients are assigned to a new drug or procedure and the other half to a placebo (or the current best option). Typically, half of the patients might get an inferior drug or treatment. We explore a method, real-time adaptive randomization (RTAR), that uses information observed up to the time of the next assignment to best allocate patients to treatments, balancing known current and unknown future outcomes—treating vs. learning. RTAR is based on a preplanned, but adaptive, assignment rule. Blinding can be maintained, so that neither the trialist nor the patient knows to which treatment the patient was assigned. During the trial, as the RTAR learns the “best” treatment, the RTAR assigns more patients to that best treatment than would a classical RCT. In two large-scale cardiovascular clinical trials, our simulations suggest that the RTAR would have saved lives while identifying the best post-trial treatment at least as well as an RCT. Some statistical measures are improved and others are worse. If endpoint rates in treatments would have changed dramatically during the trial, the RTAR would have adapted better than many other methods.