People with eczema are more likely to develop heart failure, but doctors do not know how much of that risk runs through depression, weight gain, or other treatable factors. This project builds statistical tools to answer that kind of question using the messy, irregular data found in real GP records. Existing causal mediation methods assume patients visit their doctor at regular intervals, but in practice people attend appointments sporadically and for different reasons. A patient who avoids the GP may look healthier than they are, skewing results. The researchers will develop Bayesian models that handle both irregular visit timing and the fact that some patients are more likely to seek care than others, producing unbiased estimates of how one condition leads to another. If successful, the open-source software will let epidemiologists mine the UK’s vast electronic health records for causal pathways linking dozens of conditions—eczema to heart failure, diabetes to depression, and beyond. This is methodological fundamental science: it does not test a specific drug or intervention, but it provides the analytical engine needed to turn routine clinical data into reliable evidence about disease mechanisms. Better causal estimates could eventually guide preventive care, helping clinicians know which intermediate factors to target.
View original technical description
Context It is important to understand the reasons why some people develop particular health conditions and others do not. If we can understand the reasons why people with one health condition may go on to develop other conditions, we can attempt to intervene to prevent this from happening. For example, it is known that people suffering from eczema are more likely than the general population to suffer from heart failure. Some potential reasons for this are increased body mass, or depression caused by coping with eczema, increasing the chance of later heart failure. Causal mediation analysis is a way of investigating the extent to which the impact of a particular condition on a later health outcome is mediated through an intermediate variable which is affected by the exposure and in turn affects the outcome. In the example above that means how much of the effect of eczema on later heart failure is due to depression or weight gain, and how much is for other reasons. Electronic Health Records (EHRs) are the records of medical appointments that are made when you have a health consultation. This project will use anonymised data from a network of GP practices across the UK. For each (anonymous) patient, the data set contains information on all GP visits over a period of several years, allowing us to study the development of health conditions over time (longitudinal measurements). Challenges the project will address Existing analysis methods for mediation with time-to-event outcomes assume longitudinal measurements are made on a regular time grid. Data from Electronic Health Records does not satisfy this condition, since people do not visit GPs at regular intervals. This means that using existing methods to assess mediation in EHRs may give biased results. Additionally, different patients may have different thresholds for visiting their healthcare provide despite having similar symptoms. This is called informativeness of visits, and may lead to a second form of bias if not accounted for in the analysis methods. Aims and Objectives Our project aims to solve these two challenges, by developing new statistical analysis models and software. We will use the Bayesian analysis framework, since this allows all the variables followed over time to be modelled simultaneously, and include full uncertainty on the results. In order to understand in which scenarios the new method are needed, we will carry out a systematic study to quantify the bias caused by using existing analysis methods for EHR data. Synthetic data used for model development and testing will be derived from the real EHR data, ensuring that we are building realistic and useful models. We will make our new analysis methods available through freely available, open-source software, thus enabling researchers to benefit in future analyses. Benefits of the research EHRs provide excellent opportunities for learning about health conditions. They are much more representative of the general population than small studies designed to look at one disease at a time. This research will help improve health through better understanding of disease causing mechanisms, by enabling the utilisation of these huge data sets.
Plain English summaries and category classifications on this site are generated by AI and may not perfectly reflect the original research.
Is something wrong? Let us know