POS0964 DYNAMIC PREDICTION OF PSORIATIC ARTHRITIS IN A COHORT OF PSORIASIS PATIENTS USING UK PRIMARY CARE ELECTRONIC HEALTH RECORDS.

Background: Prognostic prediction tools to identify people with psoriasis at risk of developing psoriatic arthritis (PsA) are an active area of research [1]. However, existing methods often rely on secondary care cohort studies which can lead to a small sample size and restrict model complexity. The Clinical Practice Research Datalink (CPRD) is a large, population based longitudinal cohort which presents the opportunity to develop Dynamic Prediction Models (DPMs). DPMs allow us to account for changes over time and update predictions using the most recent covariate information. One way to do this is by using landmark models, which have been used for similar purposes in other disease areas [2].

Objectives: To develop a DPM to identify predictors for the development of PsA in UK primary care.

Methods: We conducted a cohort study using routinely collected UK primary care data from the CPRD. Patients with incident psoriasis aged between 16 and 89, and no prior evidence of PsA, were identified and followed between 1998 to 2019. Relevant clinical predictors were identified using Read codes, with code list developed in consultation with rheumatologists. These included musculoskeletal symptoms, blood tests and prescriptions, building on prior work by Green [3]. Patient demographics including age, gender, BMI, alcohol, and smoking status were also included in the model. A DPM was developed using a landmark analysis framework [4]. Landmarks are time points in the study period from which we wish to make predictions, with models fitted using the subset of the patients still at risk at that time and their most recent covariate information. We used annual landmarks post psoriasis diagnosis to predict the development of PsA within 3 years. We fitted a Cox proportional hazards model, allowing a different baseline hazard for each landmark to produce a single set of variable coefficients. We report the hazard ratio and 95% confidence intervals. Variable selection was conducted using a lasso method with 5-fold cross validation to balance model complexity and predictive accuracy.

Results: We identified 122,330 incident cases of psoriasis of whom 2,460 developed PsA, with a median follow up time of 5.67 years. Hazard ratios and 95% confidence intervals are shown in Figure 1 for the demographic variables, and Figure 2 for the clinical variables. From Figure 1, we see the highest hazard ratios in age categories between 30 and 50, and the lowest in those over 60. We also see an increased risk in men, and those with BMI categories over 25. Of note, we found an increased risk of developing PsA in patients with a primary care C-reactive protein (CRP) blood test, regardless of the result of the test (Figure 2). We also found an association in those with a high plasma viscosity (PV). Psoriasis treatment, which included phototherapy or prescriptions for acitretin, ciclosporin, or dimethyl fumarate was also associated with an increased risk (HR 2.61, 95% CI 2.21 – 3.09), which could be a proxy for psoriasis severity. Patients with a non-steroidal anti-inflammatory drug (NSAID) prescription also showed an elevated risk (2.24, 2.09 – 2.39). From the musculoskeletal symptoms, we found the largest hazard ratios in patients with Read codes for unspecified arthritis (not associated with a particular joint) (2.67, 2.32 – 3.09), finger pain (2.28, 1.89 – 2.75) and knee swelling (2.22, 1.78 – 2.77).

Conclusion: We used a landmarking approach that accounts for changes over time to identify factors for the development of PsA. Future work will focus on out-of-sample predictions using alternative models that are tailored towards prediction, under the same landmarking framework.