The epidemiology of disease clusters in adults with multimorbidity

The project is supported by the University of Liverpool Doctoral Network in Future Digital Health, which is directed at creating and maintaining a community of AI health care professionals that can realise the benefits that AI can bring to Health Care. The vision is that of a world-class centre providing high-quality doctoral training within the domain of AI for Future Digital Health. Each available PhD project has been carefully co-created in collaboration with a health provider and/or a healthcare commercial interest so that the outcomes of the PhD research will be of immediate benefit. The network will be providing doctoral training, culminating in a PhD, in a collaborative environment that features, amongst other things, peer-to-peer and cohort-to-cohort based learning. On completion students will be well-placed to take up rewarding careers within the domain of AI and Digital Health.

With an increasingly elderly population, the management of patients with multimorbidity (co-existence of two or more chronic medical conditions) is the biggest challenge facing health-care systems in developed countries. The presence of multimorbidity limits life expectancy owing to frailty, recurrent hospital admissions, and polypharmacy [1]. A combination of physical and mental multimorbidity is common across the lifespan; more so in older people [2] and those from deprived backgrounds [3].
Large scale studies have defined the prevalence of multimorbidity. What remains to be addressed is the way in which diseases co-localise and the rate of accumulation of subsequent diseases states, the variation across age groups and demographics and which disease clusters are associated with worse clinical outcomes. Defining disease clusters will enable a better understanding of the antecedent factors and upstream determinants associated with progressively accumulating diseases, potentially providing new pathways for prevention and longer healthier life-span. It will also enable the design of future therapies to target multiple pathways across diseases with different aetiologies that would not ordinarily have the same treatment goals.

This project will utilise existing data available within the UK Clinical Practice Research Datalink (CPRD), which contains the electronic healthcare records (EHRs) for about 8% of the UK population. EHRs use a coding system called Read Codes that identify disease states. An individual disease (such as diabetes) will have a number of Read Codes associated with it; including those for diagnosis, monitoring, complications and so on. We will use previously validated Read Code lists that contain all the Read codes used to identify a particular diagnosis. The project will be undertaken in close collaboration with AIMES.uk, a Liverpool based cloud services provider who specialise in the deployment and use of digital technologies, such as “big data” technologies, for the transformation of health care.

The central idea is to utilise novel clustering and Machine Learning (ML) techniques to understand the way in which different diseases commonly co-exist or “cluster” together. A range of ML technologies will be used including longitudinal pattern mining and temporal clustering. The patterns in this context are sets of data attributes (Read Codes) that frequently co-occur. Algorithms exist that can be adapted to effectively and efficiently identify such patterns. The identified patterns between different diseases will enable the initial development of the ML approach to cluster commonly occurring diseases together and predict disease trajectory. Comparison to traditional cluster methods and time-to-event analyses will also be undertaken, followed by validation of the technique in the context of the whole CPRD database.

The initial pattern mining will be developed using a restricted “training” CPRD dataset and adapted such that weighted attributes (Read Codes) are considered and weighted patterns produced (some Read codes are more significant than others). The next stage will be to introduce the temporal element. The initial pattern mining can be conceptualised as two-dimensional (2D), records and attributes; the addition of time stamps adds a third dimension and consequently what can be conceptualised as 3D patterns will be required. These 3D patterns will be used to derive prediction models directed at particular conditions or combinations of conditions. A number of prediction model generators will be considered and experimented with. Once a satisfactory model has been identified the research will return to the original data set in its entirety, considering scenarios of increasing magnitude in terms of number of records and/or number of attributes.