Leiden clinical prediction rule for undifferentiated arthritis

Clinical background

In new-onset arthritis that does not meet the classification criteria for rheumatoid arthritis or another specific rheumatological diagnosis, that is, undifferentiated arthritis, the clinician faces a difficult decision. About a third of these patients develop rheumatoid arthritis within a year, while 40 to 50 per cent go into spontaneous remission [1]. Starting disease-modifying treatment (DMARD) in all of them would mean overtreating a large proportion, while withholding it risks missing the window in which early treatment has the greatest effect in preventing joint destruction.

The Leiden clinical prediction rule for undifferentiated arthritis was developed precisely to structure this decision. It combines readily available clinical and laboratory variables into a score corresponding to the probability of progression to rheumatoid arthritis within a year, and aims to identify both those whose risk is low enough for watchful waiting to be safe and those whose risk is high enough to justify early DMARD treatment.

Calculating the Leiden clinical prediction rule

The score is a weighted sum of nine variables, the weights deriving from logistic regression with progression to rheumatoid arthritis within one year as the outcome:

$$\text{Score} = 0{.}02 \times \text{age} + \text{sex} + \text{symptom distribution} + \text{symmetry} + \text{morning stiffness} + \text{tender joints} + \text{swollen joints} + \text{CRP} + \text{RF} + \text{ACPA}$$

The variables are scored as follows:

The score ranges from 0 to 14. The rule was derived from the Leiden Early Arthritis Clinic, an inception cohort at Leiden University Medical Center that since 1993 has enrolled patients with arthritis of less than two years' symptom duration. Of 1,700 patients enrolled, 570 had undifferentiated arthritis at the first visit, and it was this subgroup that was used for model development. The outcome was progression to rheumatoid arthritis by the ACR 1987 criteria within one year. The variables were selected by logistic regression, and the model was internally cross-validated to control for overfitting [1].

Interpretation in practice

The calculator divides patients into three risk bands:

In the original derivation study, each score value corresponded to a percentage probability of progression to rheumatoid arthritis [1]. External validation confirmed that a score $\geq$8 had a high positive predictive value, while a score $\leq$6 had a high negative predictive value [2].

Validation and performance

The derivation cohort. The AUC was 0.89 in the full model and 0.87 after cross-validation, indicating good discrimination and limited overfitting [1].

External validation in three European cohorts. van der Helm-van Mil et al. validated the rule in three independent cohorts from the United Kingdom, Germany and the Netherlands [2]. Since the VAS for the intensity of morning stiffness was not available in all cohorts, it was substituted by the duration of morning stiffness. The overall AUC was 0.88, with cohort-specific values of 0.83 (United Kingdom), 0.82 (Germany) and 0.95 (the Netherlands). The negative predictive value for a score $\leq$6 was 83 per cent, 83 per cent and 86 per cent respectively. The positive predictive value for a score $\geq$8 was 100 per cent, 93 per cent and 100 per cent respectively [2].

Systematic review and meta-analysis. McNally et al. identified four validation studies comprising six datasets (n = 1,084) [3]. At a threshold of $\geq$9, the pooled specificity was 0.99 (95 per cent CI 0.95–1.00) and the sensitivity 0.31 (95 per cent CI 0.24–0.37). With a pre-test probability of 40 per cent, a score of $\geq$9 raised the post-test probability to 93.6 per cent. A lower threshold of $\geq$8 gave a positive likelihood ratio of 9.5 (95 per cent CI 6.21–14.54). The authors noted, however, methodological limitations in the included studies and recommended further validation [3]. This meta-analysis therefore proposed $\geq$9 as the optimal threshold for high risk, which is somewhat higher than the calculator's $\geq$8.

Indian validation. Ghosh et al. validated the rule in a cohort of 50 patients with early symmetrical polyarthritis in Kolkata, of whom 54 per cent developed rheumatoid arthritis within a year [4]. The AUC for the Leiden score was 0.897 (SE 0.043). No patient with a score $\leq$4 progressed, and all those with a score $\geq$7 developed rheumatoid arthritis. The study was small and came from a population in which post-infectious arthritis is commoner than in European cohorts, but the results support the rule's usefulness outside western European populations as well [4].

Limitations

The rule applies only to patients with new-onset undifferentiated arthritis, that is, arthritis that does not meet the classification criteria for rheumatoid arthritis or another specific rheumatological diagnosis at the time of assessment. Patients who already meet the ACR/EULAR 2010 classification criteria for rheumatoid arthritis should not be scored with the rule, since the diagnosis has already been made and treatment decisions should be taken independently of the score.

The rule has not been validated for patients with arthritis of more than two years' duration, and the derivation cohort consisted of patients referred to a rheumatology clinic. Performance in primary care, where the prevalence of rheumatoid arthritis is lower and the mix of differential diagnoses different, is less well studied.

A practical limitation is that the intensity of morning stiffness is measured on a 100 mm VAS, which requires the patient to self-rate at the time of assessment. In the external validation this measure was missing in several cohorts and had to be replaced by the duration of morning stiffness, indicating that the variable may be difficult to standardise between clinics [2].

The indeterminate zone (score 6 to <8) is a genuine problem. In the original study these patients could not be classified reliably, and the meta-analysis confirms that no single threshold in this range provides both satisfactory sensitivity and satisfactory specificity [3]. For these patients, overall clinical assessment, possible imaging and follow-up must guide the decision.

ACPA carries twice the weight of RF (2 points versus 1), reflecting the stronger and more specific association of ACPA with progression to rheumatoid arthritis. This does, however, presuppose that ACPA testing is available. If ACPA is missing, the model loses one of its most discriminating variables and the score becomes less reliable.

References

1. van der Helm-van Mil AH, le Cessie S, van Dongen H, et al. A prediction rule for disease outcome in patients with recent-onset undifferentiated arthritis: how to guide individual treatment decisions. Arthritis Rheum. 2007;56(2):433–40. PMID: 17265478
2. van der Helm-van Mil AH, Detert J, le Cessie S, et al. Validation of a prediction rule for disease outcome in patients with recent-onset undifferentiated arthritis: moving toward individualized treatment decision-making. Arthritis Rheum. 2008;58(8):2241–7. PMID: 18668546
3. McNally E, Keogh C, Galvin R, et al. Diagnostic accuracy of a clinical prediction rule (CPR) for identifying patients with recent-onset undifferentiated arthritis who are at a high risk of developing rheumatoid arthritis: a systematic review and meta-analysis. Semin Arthritis Rheum. 2014;43(4):498–507. PMID: 24139249
4. Ghosh K, Chatterjee A, Ghosh S, et al. Validation of Leiden Score in Predicting Progression of Rheumatoid Arthritis in Undifferentiated Arthritis in Indian Population. Ann Med Health Sci Res. 2016;6(4):205–10. PMID: 28480094