The clinical relevance of a prediction rule for disease outcome in patients with undifferentiated arthritis: Comment on the article by van der Helm-van Mil et al

To the Editor:

We read with interest 2 recent articles by van der Helm-van Mil et al, in which the authors describe the development and validation of a prediction rule for disease outcome in patients with recent-onset undifferentiated arthritis (UA). The prediction rule showed excellent discriminative ability for assessing the likelihood that a patient's disease will progress to rheumatoid arthritis (RA), both in the derivation cohort and in 3 validation cohorts. Yet, we think that its relevance for clinical practice is limited, as a consequence of 2 shortcomings in the study design.

First, disease outcome was defined as fulfillment of the American College of Rheumatology (ACR; formerly the American Rheumatism Association) 1987 revised criteria for RA. This leads to incorporation bias, which results when the index test (prediction model) forms part of the reference standard (arthritis outcome). The prediction model described by van der Helm-van Mil and colleagues includes 5 of the 7 ACR criteria: morning stiffness, arthritis of hand joints, symmetric arthritis, polyarthritis, and rheumatoid factor positivity. Incorporation bias causes overestimation of the discriminative ability of the model. Another consequence of the choice of the ACR criteria as a reference standard is that the model cannot be applied to the considerable number of patients with UA and persistent/erosive arthritis who did not fulfill the ACR criteria (n = 149 [26%]) (1), even though, in daily clinical practice, these patients are treated with disease-modifying antirheumatic drugs according to the recommendations of the European League Against Rheumatism for the management of early arthritis. When disease outcome (i.e., RA) is instead defined as persistent/erosive arthritis, this problem does not occur.

Second, using UA as the diagnosis for which the prediction model is developed decreases the clinical relevance of the study, since UA is defined as arthritis not fulfilling the classification criteria of any of the rheumatic diseases. However, the diagnostic value of these criteria in early arthritis is limited or unknown in most cases. Therefore, UA is a disease entity that is ill-defined and not evidence-based. For example, the specificity of the ACR 1987 classification criteria for RA is not 100%, as is assumed in the prediction rule described by van der Helm-van Mil et al. Actually, the 1987 ACR criteria for RA have proven to be relatively poor predictors of arthritis outcome parameters. We believe it would be more useful to develop a model that can be applied to every patient with early arthritis, because in most cases (except for clear cases of [pseudo]gout that have been confirmed by the presence of crystals, and cases of bacterial arthritis that have been confirmed through cultures), the clinician is not 100% certain of the diagnosis/prognosis.