Improving the planning and monitoring of recruitment to clinical trials

Background: Successfully recruiting the pre-specified number of participants in a clinical trial remains a difficult challenge that negatively impacts all stakeholders. The approaches used to predict and monitor recruitment, including sources of information utilised, remains frequently hidden and unreported. There is an increasing number of publications describing statistical models for recruitment prediction, however there is not enough evidence about how this is done in practice. Methods: We conducted three systematic reviews to identify (1) statistical models used for recruitment prediction at the design stage of a trial, (2) methods to monitor patient recruitment and (3) statistical models used for prediction during trial conduct. To determine methods used in practice, a cohort of 125 RCTs was investigated regarding the reporting of predicted and observed recruitment. In addition, two surveys were conducted, one to statisticians working in clinical trials and the other to the chief investigators of newly funded trials. To facilitate the implementation of selected models identified, we developed an interactive web application with Shiny. Using feedback from the statisticians’ survey, a new approach building on the Poisson model is provided to address concerns around flexibility and complexity of recruitment process. Results: Existing models to predict recruitment at the design stage of the trial were either deterministic or stochastic, including Poisson, Poisson-Gamma, Bayesian and simulation models. Models were increasingly complex when used for ongoing recruitment prediction where accrual data were available. Conversely, for monitoring of patient recruitment against initial targets, the methods identified were simplistic and included tables and graphs to present the expected versus the actual number of patients recruited per month. Recruitment prediction reporting in main trial publications was often limited to stating the sample size target. The survey of chief investigators indicated that the data source most commonly used to predict trial recruitment was audit data from across multiple centres with the impact of specific eligibility criteria being the most frequently adjusted factor. The survey of statisticians indicated that statisticians are not always involved in recruitment prediction, and simple approaches are mainly used for both recruitment prediction and monitoring. The Shiny application developed bridges the gap between development and implementation of some models. The new web-based tool based on the Poisson model, which focuses greater attention on allowance factors whilst maintaining stochasticity but minimising complexity, may help investigators to better plan, monitor patient recruitment, and in decision making about the corrective actions required. Conclusion: This thesis contributes to knowledge enhancement of the methods used for recruitment prediction and monitoring of patients in clinical trials and in providing an interface to facilitate implementation. In addition, a simple model is provided, which places the emphasis on allowing for factors that reduce recruitment capacity. This work will assist investigators with choosing the right model/approach for their trial leading to improvements in the accuracy of recruitment prediction and reducing waste in research