Enhancing the design and analysis of cluster randomised trials using machine learning

Trials are important to evaluate the safety and efficacy of new treatments or interventions. One type of trial, called a cluster randomised trial (CRT) uses pre-existing groups of individuals - known as clusters - who are randomly allocated to different treatments. It means that every member of the same cluster, for instance members of the same family or patients from the same hospital, will receive the same treatment. This type of trial is very useful in real world settings where individual randomisation to treatments is not possible or the intervention is naturally applied to a whole cluster. However, this type of trial requires the use of specific analysis methods to account for the similarity in the response to treatments among members of the same cluster. Moreover, cluster randomised trials are often prone to bias, which can be corrected at the analysis stage with the use of additional information about the individuals and clusters included in the trial. This additional information can also be used to improve the precision of the trial, and to identify the individuals who will most benefit from the intervention being tested. Therefore, it is crucial to select the important variables and use appropriate statistical methods accounting for these variables to obtain an accurate estimation of the treatment efficacy and safety. However, the best way to do so remains unknown, especially in an era where there is a large amount of medical information available. In this fellowship I will draw on the emerging field of machine learning to address these methodological challenges and to determine how routinely collected medical data can be best used to improve the design of CRTs. I will use several existing trial datasets to achieve this, including CRTs in the fields of pharmacy and psychiatry, as well as data from the England Cancer Registry. Drawing on these multiple datasets but also using mathematical developments and computer-based simulation studies, I will develop and evaluate methods to improve the generalisability of CRTs, their precision, and bring us one step closer to a more personalized medicine approach for patients.

Pragmatic trials emerged in response to concerns that clinical trials, traditionally designed to assess efficacy, were failing to inform clinical practice. Assessing benefits and harms in highly selected patient populations under well-trained experienced clinical teams can lead to over-estimation of benefit and underestimation of harm in practice. Pragmatic trials, conversely, aim to assess benefits, harms, and cost-effectiveness in real-world settings, and to identify subgroups for whom the intervention is most effective. Cluster Randomised Trials (CRTs) have emerged as an important design for pragmatic trials. However, there are a number of methodological challenges that must be addressed in order for CRTs to fulfil their promise, including enhancing their generalizability, reducing bias, improving precision, and the identification of individualized intervention effects. In this fellowship I will gain theoretical and practical training in the emerging area of machine learning (ML) to address the above challenges. To achieve this, I will:
1. Develop a framework to emulate CRTs from observational studies to evaluate the effect of cluster-level interventions and compare parametric and non-parametric methods, including ML approaches, to estimate inverse-probability-weights to address confounding;
2. Provide practical guidance for researchers on how to exploit ML to improve the selection and adjustment for covariates in CRTs, in order to both increase precision by reducing the intra-cluster correlation and to minimise bias by adjusting for confounding;
3. Propose and evaluate ML methods to study heterogeneity in intervention effects in CRTs and estimate subgroup effects while maintaining valid confidence intervals.
These objectives will be achieved by developing, extending and assessing ML methods using a mixture of theory, simulation studies and applications to case studies, including CRTs in pharmacy and psychiatry and observational data from Cancer Registries.

During this fellowship, I will develop novel statistical methodology using a variety of methods, including machine learning (ML) to enhance the design and analysis of cluster randomised trials (CRTs). This research will focus in particular on reducing bias and increasing precision of CRTs, identifying subgroup of patients benefitting the most from the intervention and using external routinely collected data to better inform the design of CRTs.

Immediate beneficiaries of the project include researchers involved in planning and running CRTs, as well as statisticians with an interest in the methodology of CRTs. This research will provide tools and information which will enable researchers to harness the potential of ML to improve the efficiency of their trials.

Patients would also be major beneficiaries of this research. Designing more efficient CRTs and gaining statistical power with the use of ML algorithms can lead to trials conducted on smaller sample sizes, meaning that fewer patients would be exposed to potentially harmful interventions. Furthermore, smaller trials are usually shorter, which would reduce the delay between the start of the CRT and the moment the treatment or intervention studied is provided to the patients in practice. As such, patients would be treated more quickly. Smaller and shorter trials are also less costly, which will also be beneficial to trial funders, such as research councils and charities. My third objective, which focuses on the identification of subgroup effects, will be a step towards a more personalised medicine. Therefore, my research could allow patients to receive the optimal treatment based on their individual characteristics.

Through the development of a new framework for the emulation of CRTs from observational data, this work will take advantage of the wealth of medical information routinely collected to evaluate the effect of interventions in real-life settings, making use of existing and usually large data rather than collecting new information. This will allow researchers to obtain accurate results in a more timely manner but also replace expensive pilot CRTs usually conducted before the start of a larger-scale CRT.