Background/Purpose: Psoriatic arthritis (PsA) is a heterogeneous chronic inflammatory disease that affects the skin, joints, and soft tissues. No metrics exist to indicate whether an untreated patient is unlikely to respond to oral therapies and will require therapy with a biologic.  The objective of this study was to predict TNFi prescription among treatment naïve patients initiating methotrexate. We further aimed to identify subgroups of patients at baseline that would likely need a TNFi.

Methods: Data from the Tight Control of Psoriatic Arthritis (TICOPA) trial was used. In TiCOPA, patients with at least one tender and swollen joint who were treatment naïve were randomized into one of two arms: intensive management and standard of care.  We used all available baseline data to build models to predict the need for a subsequent TNFi within the 48 week study. We first applied a standard prediction modeling approach using stepwise Cox regression models.  However, because of correlation among baseline variables (i.e., collinearity) and instability of the models, we decided to apply novel “machine learning” methods (e.g., tree-based classification, clustering methods, and latent class analysis, LCA) using R 3.4.4 with imported packages. Target categorical variables were cut from continuous variables after graphical exploration and trial analyses.

Results: Among the 188 participants who agreed to share data for additional studies, 44 initiated a TNFi during the 48-week study period.  Exploration with comparison tests, stepwise regression, and tree-based models identified the PsAQoL score and the Patient Global Assessment (PtGA) as valuable predictors of TNFi prescription. Using LCA, we were able to define a two-class model (Figure 1A) that classified patients who were more and less likely to require a TNFi.  In addition to receipt of a TNFi, these classes were defined by a CRP ≥ 25, PsAQoL ≥ 7, and TJC ≥ 10 (the most predictive and non-redundant set of variables). We termed the classes “Less Inflamed” and “More Inflamed” for simplicity. Importantly, PtGA was a strong predictor of being in one of these classes; as PtGA increased, the likelihood of being the “more inflamed” class increased. Dividing PtGA into Low (≤40), middle (40-70), and high (≥70) categories accurately predicted class membership (Figure 1B).  In summary, PtGA was single best predictor of requiring a TNFi; baseline PtGA≥70 has a HR of 2.62 (1.42-4.80) for requiring a TNFi and correctly classifies 76% of patients.

Conclusion: At baseline, participants in the TICOPA trial could be classified into two groups, “more inflamed” and “less inflamed” based on three variables and patient global was the single best predictor of class and receipt of a TNFi. The baseline patient global assessment of treatment naïve patients with PsA yields valuable prognostic information and is an important tool for clinical decision-making in this population.