Predicting psoriatic arthritis in new-onset psoriasis: development of multivariable prediction models from an inception cohort study.

Abstract

Psoriatic arthritis is underdiagnosed in people with psoriasis. Delayed psoriatic arthritis diagnosis is associated with worse outcomes, whereas early intervention improves prognosis. Therefore, we aimed to develop and internally validate two sets of prediction models. The first set intended to identify participants with new-onset psoriasis suitable for rheumatologist referral; the second set intended to identify participants with subclinical psoriatic arthritis at high risk of developing clinical psoriatic arthritis. We also aimed to identify psoriatic arthritis predictors. We analysed data from the Stockholm Psoriasis Cohort, a Swedish psoriasis inception cohort that enrolled participants with psoriasis within 1 year of their first psoriasis lesion on non-hairy skin between Jan 5, 2001, and Dec 14, 2005. Predictors were obtained at enrolment. We developed two referral models (one with biomarkers and one without) and two models for subclinical psoriatic arthritis (one with a 3-year horizon and one with a 15-year horizon). We used recursive partitioning and penalised regression for prediction modelling and assessed the models using discrimination, calibration, and net benefit over a range of clinically reasonable risk thresholds informed by a clinician survey. People with lived experience of psoriasis were involved in study design. Of 628 participants who reported no psoriatic arthritis at enrolment (347 [55%] females, 281 [45%] males, median age 40·9 years [IQR 30·1-56·0]), 83 (13%) had concomitant psoriatic arthritis. Recursive partitioning with laboratory biomarkers stratified participants into four groups using five factors (arthralgia, fatigue, psoriasis phenotype, high-sensitivity C-reactive protein, and psoriasis disease activity) with risks of concomitant psoriatic arthritis from 1% to 62%. The penalised regression model with a 15-year horizon included nine variables and the model with 3-year horizon a subset of these. All models showed good discrimination (optimism-adjusted area under the curve: 0·76-0·84). The models showed reasonable calibration and, based on thresholds derived from a clinician survey, positive net benefit for decisions on referral and monitoring, they but appeared to be of limited value for decisions about preventive treatment. All models, but most notably the prognostic models, exhibited instability. Pain, HLA-B27, and systemic inflammation were the strongest predictors for future psoriatic arthritis. Predictors of psoriatic arthritis are already present at psoriasis onset. The models developed here could support clinical decision making, but they require external validation before implementation. Group for Research and Assessment of Psoriasis and Psoriatic Arthritis, Hudfonden, the National Psoriasis Foundation Psoriasisförbundet, the Stockholm County Council, and the Swedish Medical Research Council.