Forecasting patient recruitment for time-to-event analysis in National Lung Matrix Trial

Introduction: The National Lung Matrix Trial is a multi-drug,
genetic-marker-directed, non-comparative, multi-centre, multiarm
phase II trial of non-small cell lung cancer. Each arm
studies the efficacy of a targeted drug in a patient population
stratified by pre-specified genetic markers using a Bayesian adaptive
design. Biomarker prevalence is a source of uncertainty so we
sought to model patient recruitment to inform analysis and
drug supply schedules. Some arms use the time-to-event (TTE)
primary outcome, progression-free survival time (PFS) and the
statistical characteristics of the design vary with recruitment
rate. Recruitment was an important source of uncertainty to
understand.
Methods: Estimated prevalence rates are used to predict average
recruitment over time and simulation is used to understand the
variability about this mean. Thall et al [2005] describe Bayesian
methods for analysing TTE outcomes in trials with interim analyses
that stop if the treatment does not offer significant benefit. We
implemented this design with a stop-go decision at interim and
final analyses. The chances of reaching the correct conclusion
(the operating characteristics) varied with recruitment rate. Using
simulation, we calculated operating characteristics under different
survival and recruitment scenarios and weighted these scenarios
together. Different weight schemes over a wide range of scenarios
allow us to check how our method would perform when patients
arrive much faster or slower than anticipated and verify the
robustness of our design.
Results: Trial operating characteristics improve with slow recruitment
because the trial accrues more information to inform a
statistical decision.
Conclusion: Forecasting patient recruitment allows trials units and
treatment centres to predict their workloads and pharmaceutical
companies to predict drug supply needs. A design whose
performance varies with recruitment rate should model this
important source of variability. We verified that our statistical design
to analyse TTE data would perform well under the most extreme
plausible recruitment scenarios.