MRC Integrative Epidemiology Unit (IEU) Seminar Series

Title: Where multi-armed bandit models meet response-adaptive randomisation for clinical trials

Abstract: Multi-armed bandit problems are an important and well-known class of models for studying the learning-earning trade-off within reinforcement learning, operations research and several other fields. Although many algorithms for these problems are theoretically well-understood and commonly used in real applications (e.g. web advertising), others remain (still) largely unused in practice. Specifically, this is the case for the extensive and rich body of literature accumulated on formulating and solving theoretical bandit models for clinical trials in which some measure of its outcome is optimized in a Bayesian setting. In such a context, the learning-earning trade-off exists between two competing goals: (1) to correctly identify the best treatment (learning) and (2) to treat as many patients as effectively as possible during the trial and after (earning). In this talk, I review some important results from this literature, present their advantages and limitations to their use for designing clinical trials in practice and discuss how bandit algorithms can be viewed (and modified) to define response-adaptive randomisation procedures. I will illustrate how different bandit algorithms perform in terms of different operating characteristics in the context of real clinical trials. I will also discuss how this line of work has resulted from an attempt of bringing theory closer to practice while providing an answer to the current challenges of therapy development for rare diseases and personalised medicine. I will also briefly discuss the potential of other applications of these results to optimizing the development of health apps and to designing surgical trials.

Biography:  Sofia is a senior statistician in the clinical trials methodology group at the MRC Biostatistics Unit (University of Cambridge School of Clinical Medicine).  She leads a team of statisticians at Papworth Trials Unit Collaboration. Sofia was awarded the first Biometrika post-doctoral research fellowship in 2014. Before that she was in the Department of Mathematics and Statistics of Lancaster University, yet working from the MRC Biostatistics Unit, as part of a joint research project on the design of multi-arm multi-stage clinical trials. Sofia did her PhD thesis on designing tractable and nearly optimal solutions to computationally infeasible constrained sequential estimation problems. 

All welcome