Patients with the skin disease psoriasis are at increased risk of developing a chronic, progressive inflammatory arthritic disease called psoriatic arthritis (PsA). In this thesis, our aim was to identify patterns in primary care clinical symptoms, prescriptions, and blood tests to identify psoriasis patients at increased risk of PsA. Statistical models were developed using longitudinal data from a large cohort of psoriasis patients identified in the Clinical Practice Research Datalink, a database of UK primary care electronic health records. Our statistical approach was centred around the use of Bayesian networks (BNs), a class of interpretable models which can capture complex relationships between variables and visualise statistical dependencies with a directed acyclic graph. Our analysis of the psoriasis cohort included both a static and dynamic approach to statistical modelling. In the static approach, we developed a BN to classify psoriasis patients based on whether they developed PsA during the study period, using clinical variables measured prior to diagnosis. We produced a model which could be used to aid in PsA screening by identifying patients at increased risk of having PsA, and we identified important primary care clinical variables associated with undiagnosed PsA. In the dynamic approach, we produced a dynamic clinical prediction model to estimate psoriasis patients’ 1-year risk of developing PsA. We developed a novel statistical methodology for producing dynamic v-year risk predictions using a BN in a landmarking framework. Landmarking is a method for producing dynamic predictions by pre-defining a series of time points (landmarks) throughout the study period at which we want to update risk predictions, making use of any updated clinical variables. The most up-to-date variable information for psoriasis patients still at risk at each landmark is identified, and whether the patients develop PsA before the end of the v-year prediction window is determined. This study design is affected by patients who were lost to follow-up before the end of the prediction window, resulting in data which is right-censored. BNs are not naturally suited to time-to-event data, and we developed a novel methodology for learning the structure of a BN from right-censored event data. The results of the analysis presented in this thesis could be used to inform clinical practice of patients at increased risk of PsA, facilitating the earlier diagnosis of the disease and improving patient outcomes. The novel statistical methodology can be used to produce interpretable dynamic risk prediction models capable of modelling complex interactions between variables, which is not limited to PsA research and could be utilised for similar tasks in other disease areas.
- Alternative Format
- Psoriatic arthritis
- Psoriasis
- Survival analysis
- Bayesian network
- Clinical prediction model
Predicting Psoriatic Arthritis: Dynamic modelling of primary care health records for earlier diagnosis of psoriatic arthritis
Rudge, A. (Author). 8 Oct 2025
Student thesis: Doctoral Thesis › PhD