Abstract
BACKGROUND: Rheumatoid arthritis (RA), particularly seronegative disease, is difficult to diagnose early, which can delay treatment initiation. This study aims to develop a binary DNA methylation (DNAm)-based algorithm to diagnose RA. METHODS: Three datasets (discovery, training, holdout) were constructed from DNAm profiles from 1366 persons (treatment-naïve RA, other inflammatory/autoimmune diseases, healthy individuals). DNAm features that differentiate RA from other inflammatory/autoimmune diseases and healthy individuals were identified using the discovery set. Our classification algorithm was developed using machine learning techniques in the training set. Its diagnostic performance, with and without serological status, was evaluated in the holdout set containing RA cases (15 seropositive, 6 seronegative) and controls (14 other arthritides, 11 healthy individuals). RESULTS: Our algorithm included 391 DNAm features. Combined with serological status, it classified RA from controls in the holdout set with the following performance: sensitivity 0.90 [95% CI: 0.70-0.99], specificity 0.88 [95% CI: 0.69-0.97], and AUC 0.96 [95% CI: 0.91-1.00]. Its performance in classifying patients with seronegative RA versus those with other arthritides was: sensitivity 0.83 [95% CI: 0.36-1.00], specificity 0.79 [95% CI: 0.49-0.95], and AUC 0.81 [95% CI: 0.61-1.00]. CONCLUSIONS: The present DNAm-based classification algorithm may be clinically useful for the early diagnosis of RA, especially in seronegative patients, which currently often poses a diagnostic challenge.