Arguing for Adaptive Clinical Trials in Sepsis

Abstract

Sepsis is life-threatening organ dysfunction due to dysregulated response to infection. Patients with sepsis exhibit wide heterogeneity stemming from genetic, molecular, and clinical factors as well as differences in pathogens, creating challenges for the development of effective treatments. Several gaps in knowledge also contribute: (i) biomarkers that identify patients likely to benefit from specific treatments are unknown; (ii) therapeutic dose and duration is often poorly understood; and (iii) short-term mortality, a common outcome measure, is frequently criticized for being insensitive. To date, the majority of sepsis trials use traditional design features, and have largely failed to identify new treatments with incremental benefit over standard of care. Traditional trials are also frequently conducted as part of a drug evaluation process that is segmented into several phases, each requiring separate trials, with a long time delay from inception through design and execution to incorporation of results into clinical practice. By contrast, adaptive clinical trial designs facilitate the evaluation of several candidate treatments simultaneously, learn from emergent discoveries during the course of the trial, and can be structured efficiently to lead to more timely conclusions compared to traditional trial designs. Adoption of new treatments in clinical practice can be accelerated if these trials are incorporated in electronic health records as part of a learning health system. In this review, we discuss challenges in the evaluation of treatments for sepsis, and explore potential benefits and weaknesses of recent advances in adaptive trial methodologies to address these challenges.

Keywords: Bayesian statistics; adaptive clinical trials; platform trials; response adaptive randomization; sepsis.

Figure 1

Schematic representation of a hypothetical adaptive platform trial. An initial block of patients is stratified based on known or candidate predictive biomarkers, and then randomized to an experimental or control arm. Once a predefined number of patients is enrolled, outcomes are observed and the data are input to the Bayesian statistical model by arm and stratum, which is used to calculate the predictive probabilities (PP) that each experimental arm will be superior to control in the final analysis. These PP are checked against predefined decision boundaries established so that arms with poor probability of success are dropped, and arms with high probability of success “graduate” to the next phase of testing. Arms with PP that do not require dropping or graduation continue enrolling subjects; arms that are removed may be replaced by new experimental treatments, accrual permitting. Finally, the PP are used to update randomization probabilities used for the next block of patients to be enrolled, and the feedback loop begins anew.