HOD2: Instrumental Variable approaches for estimating heterogeneity of treatment effects to inform personalisation using electronic health records

Personalised medicine aims to provide the right treatment for the right patient at the right time. The evidence to inform personalised medicine can come from patient's electronic health records. However, there are several problems with using these data to compare outcomes for groups of patients who receive alternative treatments. First, the patients in the two groups may differ according to characteristics that are not measured (for example, the patient's frailty). Second, the extent to which an intervention leads to an improvement in the patients' outcome (or harm) may differ according to these unmeasured characteristics. Third, it is often unknown which patient groups benefit from which interventions. Methods for addressing these three problems are currently unavailable.

This project will develop new methods that resolve these problems and provide more accurate estimates of the effectiveness and harms of new treatments that apply to individual patients. In particular, we will develop new methods that make more realistic assumptions. Instead of assuming that we know which patient groups benefit from which treatment, we will develop approaches that can learn from the data about which subgroups benefit from which treatment.

We will examine how well these new methods work in practice by testing them as part of new studies. One of these studies will examine which patients benefit from emergency surgery for common acute conditions (e.g. appendicitis), and uses data from 1.5 million hospital episodes. The other study considers which patients with type 2 Diabetes Mellitus benefit from new, more costly oral treatments, and uses prescription data from General Practice databases for 25,000 patients. Our new methods will enable us to provide more accurate, relevant evidence about which interventions work best for which patients with these two conditions. We will provide a general framework for these methods that can be applied across many different disease areas and countries.

To help future studies, we will provide tutorials and guidance on using and adapting these methods in different contexts. We will run short courses and workshops to assist those designing, analysing and interpreting studies that use electronic health records to inform treatment decisions for individual patients.

Personalisation medicine requires evidence on how the comparative effectiveness and harms of alternative treatments differs across individual patients, a concept known as Heterogeneity of Treatment Effect (HTE). Electronic Health Records (EHRs) have the potential to provide this information, but major methodological challenges must be addressed. The three major interrelated issues are that: the comparison groups may differ according to unmeasured prognostic characteristics (e.g. frailty), unmeasured confounders may also predict HTEs, and it is unknown which observed characteristics truly predict HTE.

Instrumental Variable (IV) methods can address confounding due to unobserved as well as observed characteristics. However, when there is a HTE, standard IV approaches do not provide treatment effects that can be easily interpreted for treatment decisions. Local Instrumental Variable (LIV) methods have been developed that use continuous IV to provide treatment effect estimates of direct relevance for personalisation. However, current LIV methods assume that the analyst chooses those covariates which are true treatment effect modifiers. While data-adaptive approaches to covariate selection have been developed elsewhere, they have not been extended to the selection of covariate by treatment interaction terms for IV designs.

We will develop data-adaptive IV methods that address concerns about whether the underlying parametric models are correctly specified, and selects those observed covariates which do actually modify the treatment effect. We will subject these IV methods to careful testing within simulation studies, grounded in two live NIHR-funded studies which aim to inform personalised treatment strategies. We will develop an analytical strategy for future studies. We will target translation activity at those designing, analysing and interpreting studies that aim to provide the requisite evidence to personalise clinical guidelines and treatment decisions.

The proposed development and evaluation of new data-adaptive Instrumental Variable (IV) methods that can use Electronic Health Records (EHRs) to inform personalised medicine, tackles a general problems in the use of observational data for comparative effectiveness research.

The research will benefit those designing, conducting and analysing observational data, regulators, reimbursement agencies and clinicians making treatment and funding decisions, and ultimately help lead to improved population health.

We will develop and test methods data-adaptive instrumental variable (IV) methods that can address unmeasured confounding, fully recognise heterogeneity and are less reliant on choosing the correct model specification than currently available methods. We will develop software tools and guidance to facilitate the use of these advanced methods, and provide training in their usage. Studies will be able to harness EHR to provide the requisite evidence to personalise clinical guidelines and treatment decisions. The research will help future studies make transparent, more realistic assumptions about IV methods, and provide estimates of direct relevance to inform personalisation strategies.

Our methods will also be more broadly applicable than studies that use EHR, and for example will apply to studies using clinical databases, or Biobank data where there may be many potential treatment effect modifiers. This project will benefit applied researchers working with observational data, and methodologists with an interest in advanced methods for causal inference.

The outputs from this research will benefit, pharmaceutical companies, academic researchers in health and social science, data holders (e.g. Clinical Practice Research Datalink), providing electronic health record data to researchers interested in questions of comparative effectiveness; policy makers (e.g. NICE, DoH), and research funding bodies (e.g. NIHR).