Evaluating effects of complex treatments using large observational datasets: from population to person

Studies of the effects of treatments on health outcomes are the basis of decision-making in health care. It is increasingly recognised that observational data from sources such as disease registries and electronic health records can provide 'real world' evidence about effects of treatments. The advantage of observational data is that they contain large numbers of individuals and a more diverse set of individuals taking a given treatment than would be included in a randomised trial. A major challenge is that studies of treatment effects based on observational data are prone to bias, including due to those receiving a given treatment tending to be more ill than those who do not, and therefore being at higher risk of adverse health outcomes (i.e. confounding). Statistical methods have been developed in recent years that enable use of observational data to estimate effects of treatments while avoiding such biases, and these are referred to as causal inference methods. Most studies of treatment effects focus on average effects for the patient population. This research programme focuses instead on developing methods that will allow us to provide more personalised information about what an individual's health outcomes would be expected to be under different treatment choices, by using observational data collected over time. We will use these methods to investigate questions about effects of treatments used in four health areas.

There are four work packages in this project. The first focuses on development of innovative statistical methods for 'counterfactual prediction', which refers to making predictions about what a person's risk of an outcome would be under different treatment choices given their individual characteristics such as their age and health status. We will also show how to assess how accurate counterfactual predictions are, which is vital if they are to be used in practice. The second work package focuses on methods for answering the related question of when it may be best to start a particular treatment, in terms of when a clinical measurement reaches a given level, which is referred to as a dynamic treatment strategies.

The methods will be applied to tackle questions about treatment effects in cystic fibrosis (CF) (work package III) and in hypertension, diabetes, and intensive care medicine (work package IV). CF is the most common inherited disease in the UK, and recent years have seen the introduction of precision medicines for a large proportion of the UK CF population. This raises important questions about the benefits of prior existing treatments, for which patients, and when. We will use data from the UK CF Registry to develop counterfactual predictions of outcomes under different treatment choices, and to investigate dynamic treatment strategies. Through three further projects we will apply the methods in other health areas using large-scale electronic health records data sets. These will include using GP records data from the Clinical Practice Research Datalink to investigate (1) the optimal timing of initiation of blood pressure medications for preventing cardiovascular disease and whether this depends on other patient characteristics, and (2) how patient characteristics and timing of treatment start impact on benefits of second-line treatments for people with type 2 diabetes who have been prescribed Metformin. We will also use high-frequency data on patients admitted to University College London Hospitals to develop personalised predictions of risk of death for post-operative patients under the options of admitting the patient to intensive care or not.

In summary, this project will develop methods that allow us to learn about what the impact of different treatment decisions could be for different people. This will ultimately result in information that allows patients and their doctors to make more personalised decisions about their treatment options.