Statisticians at Cytel are staunch advocates of the use of Bayesian methods in clinical trials. This summer's Joint Statistical Meeting will feature a panel on "Study Milestone Timeline Projection in Multi-Center Trials: Bayesian and Frequentist Approaches with Real-Life Applications" with three papers co-authored by Cytel experts. The Cytel papers will center on Bayesian approaches for enrollment prediction, which is essential for proper trial forecasting and interim decision-making.
During the Joint Statistical Meeting next week, Cytel will also offer a Computer Technology Workshop on how software can make Bayesian methods easier to apply to dose-escalation.
According to Cytel Consultant Pantelis Vlachos, one of the key benefits of applying Bayesian methods to clinical trials is the continual updating of inferences for purposes of predictive probabilities. For example, Bayesian methods can determine how a current cohort of patients will respond to treatment based on data gathered during the earlier course of a trial. As the video below demonstrates, this leads to more informed patient care and more precise dose-escalation.
Abstracts of JSM Papers
Bayesian Enrollment Timeline Projection
The accuracy of enrollment timeline projection plays an important role in planning and execution of clinical trial operations. This is particularly critical in time to event trials, where accuracy of event timeline projection is contingent on corresponding enrollment prediction. This talk will discuss a simulation approach for enrollment timeline projection using Bayesian Poisson-Gamma model in two real-life multi-center oncology trials. The model parameters include site structure of a typical clinical trial in terms of countries, sites, activation plan, and enrollment rates/caps at country/site level. The prior shape and scale parameters for Gamma distribution are estimated on the basis of the specified enrollment rates. During the trial, all available information on observed enrollments and site activations is progressively used for Bayesian updates of the Gamma parameters. The model has been validated using past data from several real-life clinical trials. In addition to the enrollment timeline projection in terms of 95% credible intervals and other specified percentiles of interest, the model also generates probability distributions of reaching specific enrollment milestones.
Bayesian Clinical Events Timeline Projection
Reliable events timeline projection is critical to support decision-making in planning and execution of time-to-event clinical trials. This talk will discuss a simulation approach for events timeline projection using Poisson-Gamma model in two real-life multicentre oncology trials. The model works in conjunction with related enrollment timeline projection model. The model parameters include hazard rates for piecewise exponential distribution for clinical events and rates of dropout. The parameters are estimated on the basis of design assumptions specified in a typical clinical trial protocol. The parameters are used to estimate shape and scale parameters of related Gamma distributions corresponding to each segment of piecewise exponential distribution. During the trial, all available information on observed enrollments, and clinical events/dropouts is progressively used for Bayesian updates of the Gamma parameters. In addition to the events, timeline projection in terms of 95% credible intervals and other specified percentiles of interest, the model also generates probability distribution of reaching specific events milestones, including effective scheduling of DMC review meetings.
Sample Size Re-Estimation in Multi-Center Oncology Trials with Time-to-Event Endpoint Using Bayesian Enrollment Timeline Projection and Clinical Event Prediction
In a time to event study, the defining element of accrual goal is to achieve the target number of events, rather than a planned sample size. It is known that sample size re-estimating based on revised estimate of event rate from blinded analysis of aggregate study data enhances efficiency with limited risk of introducing bias/impairing interpretability (FDA Guidance for Industry, 2010). Cook (2003) used Markov Chain method to estimate the probability for a subject to be in event state or treatment state at a given survival time point using either parametric (Weibull) or empirical per-time point hazard rates. These probabilities are combined with actual and future enrollments to generate the number of subjects in event state at a given "study" time point thus guiding the sample size modification decision. We propose to use Bayesian Poisson-Gamma model to predict future enrollments (Nitin et al 2014) in applying above method. An application example using actual data will be presented with the comparison/discussion of using parametric vs. empirical per time point hazard rates in the algorithm. The extension of this method including the lost to follow up state will be also discussed.