Statistical challenges in utilising Electronic Health Records for medical research

Missing data in electronic health record research: challenges and (some) solutions

Studies conducted using routinely collected data - such as electronic health records (EHR) - often suffer from missing data.  While the field of missing data methodology is comparatively well developed, certain features of EHR data mean that the application of missing data methods requires specific consideration in this setting. One key issue is the sheer size of the available data, with large numbers of patients, but also large amounts of data collected on each patient over many years. A second issue is that data were not collected for the purpose of research, leading to data-dependent sampling: measurements are often only taken when the patient visits their GP, an action which is often prompted by reasons related to the health status of the patient at that time.

This talk will explore problems of missing data in studies using EHR data, paying particular attention to features specific to this setting. A range of possible missing data methods will be discussed, and illustrated on a range of real EHR studies.

Linkage matching errors and quantifying uncertainty for model fitting

The talk will describe how the inherent uncertainty about the correctness of record matches can be quantified and utilised in subsequent statistical modelling. The proposal is to extend a Bayesian missing data algorithm to incorporate data priors estimated from a particular record matching algorithm. Implications of this approach for the development of population record 'spines' will be mentioned.

Estimating cardiovascular disease risk in electronic health records with incomplete records and repeated measurements of risk predictors

Stratification of individuals according to their estimated cardiovascular disease (CVD) risk is used to guide clinical decision-making. Current UK guidelines for CVD risk assessment recommend the use of already recorded risk factors in electronic primary care records to prioritise patients for a full formal risk assessment, although there is no guidance on how this should be achieved.

We present a computationally feasible statistical approach to address the methodological challenges in utilizing historical repeat risk factors measures recorded in primary care records to systematically identify patients at high risk of future CVD disease. The approach is principally based on a dynamic two-stage landmark model. The first stage estimates predicted current risk factor values from all available historical repeat risk factor measurements by landmark-age-specific multivariate linear mixed-effects models with correlated random-intercepts, which account for sporadically recorded repeat measures, unobserved data and measurements errors. The second stage predicts future disease risk from a sex-stratified Cox proportional hazards model, with predicted current risk factor values estimated from the first stage.

We have developed and internally validated a dynamic 10-year cardiovascular disease risk prediction model using electronic primary care records for age, diabetes status, hypertension treatment, smoking status, systolic blood pressure, total and high-density lipoprotein cholesterol from ~2 million individuals in ~400 primary care practices in England and Wales contributing to Clinical Practice Research Datalink. We propose using such models as pre-screening tools for identifying individuals who may be at greatest health need of a more formal CVD assessment. Using public health modelling, we identify optimal pre-screening risk thresholds for inviting individuals in for a formal risk assessment.