2B-02 - Application of Bayesian Detection of potential Risk using Inference on Blinded Safety data (BDRIBS)

Background: During drug development, proactive monitoring of safety information and contextualization utilizing epidemiological data are critical components of safety monitoring plans. The 2021 FDA Draft Guidance on Safety Reporting Requirements and Safety Assessment for Investigational New Drug Studies requires periodic review of safety data. As an alternative to unblinded safety monitoring, the FDA recommends a blinded approach to minimize risks to trial integrity.

Objectives: The main objectives are to evaluate the performance of the Bayesian Detection of potential Risk using Inference on Blinded Safety data (BDRIBS) as a method for blinded safety monitoring and highlight the role of epidemiology in contextualizing safety data based on the application of BDRIBS through simulation.

Methods: BDRIBS was used to estimate the posterior distribution of the relative risk (treatment vs control) of a pre-selected adverse event (AE). As BDRIBS requires the use of rates (i.e., person-time), epidemiologic data on the incidence rate of the AE in the disease population (i.e. background rate) was used as an approximation to the incidence rate in the trial control arm. Considerations for selecting a suitable background rate included appropriate identification of the AE, data source, and trial population. Simulation studies were conducted to evaluate the performance of BDRIBS using conventional unblinded statistical analysis as a benchmark to compare the operating characteristics (OCs) and understand the impact of sample size, misspecification of background rates, and non-constant hazards.

Results: In scenarios where the background rate exactly matched the true incidence rate in the control arm, BDRIBS demonstrated excellent OCs, with negligible difference in the sensitivity and specificity compared to the conventional unblinding approach. As expected, the optimum OCs were observed in scenarios with larger sample sizes, non-rare AEs, and with large true relative risk. In scenarios with misspecified rates, non-constant hazards, and very rare AEs ( < 1.5%), OCs were sub-optimal especially when background rates were overestimated and, to a lesser extent, when hazards were not constant. For very low incidence rates, the sensitivity for BDRIBS to detect a safety risk was comparable to the unblinded conventional approach when the rate was correctly specified.

Conclusions: BDRIBS exhibits favorable OCs in most settings where key assumptions are met, including accurate background rates and constant hazards. OCs may become sub-optimal when these are violated. Reviewing and contextualizing epidemiological data and conducting simulations are key to supporting a thorough evaluation on whether BDRIBS is fit for purpose.