Investigating comorbidities with causal networks

The problem of comorbidities in clinical research is huge - especially in older populations, once we begin looking at diseases such as cardiovascular disease and diabetes, non-ignorable numbers of patients have multiple conditions, and all of them have a whole variety of risk factors that may interact in various ways. The biological mechanisms underlying this are highly complex, and even with the vast amounts of data currently available, we have limited approaches to comprehending it in its totality. Causal inference, especially causal network analysis, is the key piece of Biostatistics and Data Science that can use the influx of data to drive these advances and gain further insights into different and/or overlapping pathways to multiple health outcomes.

Tackling all those causal links at once is an excessively challenging task. Therefore, we will be adapting pre-trained natural language processing algorithms (e.g. Mikolov et al 2013) to objectively select the disease outcome associated factors. We will then focus on genetic data, by developing a polygenic risk score (Bond et al 2022) that is predictive of the selected risk factors. Next, we will bring in the principles of Mendelian randomization by using the polygenic risk score as the instrument and Bayesian networks (Kelly et al 2022) to leverage data with large numbers of exposures and potential comorbidities to build causal networks. These causal networks will be used to help make informed decision on selecting a few important risk factors that are easy to manage and intervene on.

In particular, we will look into stroke, hypertension and diabetes using the UK Biobank data (access to be applied by student once the project starts). We shall use these networks to investigate the underlying causal mechanisms when it comes to comorbidities, as these may be difficult to ascertain when the directionality of causes and effects and interacting pathways are concerned.