Using a genetic probability tool (g-prob) to aid diagnosis of patients with inflammatory arthritis: results from the norfolk arthritis register

Abstract Background/Aims Recent translational advances in genetics report the ability to accurately predict the diagnosis of patients presenting with synovitis, providing potential to accelerate treatment and improve patient outcomes. One tool is G-PROB, which uses genetic information to calculate conditional probabilities, known as G-probabilities ranging from 0 to 100%, for defined diseases. In the original study, G-PROB was configured to discriminate between patients with rheumatoid arthritis (RA), systemic lupus erythematosus (SLE), psoriatic arthritis (PsA), spondyloarthritis (SpA), gout and “other rheumatological diseases” using reported bias-adjusted odds ratios from 250 known single nucleotide polymorphisms and human leukocyte antigen variants of uncorrelated risk variants. The original study tested G-PROB on 243 patients with synovitis. Our aim was to assess whether G-PROB could aid diagnosis using data from the Norfolk Arthritis Register (NOAR), a large observational cohort of patients with early inflammatory arthritis. Methods Genotypes, and clinician diagnosis were obtained from NOAR. The same prevalence settings and risk variants as the original study were used. Six G-probabilities each corresponding to one disease were created for each patient and performance was assessed using linear regression without intercept, negative and positive predictive values (NPV, PPV), and receiver-operator-curve (ROC) analysis. Results From NOAR, 2031 genotyped patients were identified and underwent case note review to determine the clinician diagnosis. Clinician diagnoses included RA (n = 767), PsA (n = 106), SpA (n = 15), SLE (n = 14), gout (n = 5), and “other” (n = 65). For n = 1059, case notes were not available resulting in exclusion. The mean G-probability was 41% for those which corresponded to clinician-defined disease, which was significantly higher compared to 12% for those that did not (95%CI -0.30 to -0.28). As reported in the original study, G-probabilities were concordant with real disease status (β regression coefficient of 1.03 vs 0.99, where 1.00 is ideal). We found 42% of all G-probabilities were <5%, corresponding to a NPV of 99%, where it was possible to deprioritise >1 disease for 100% of patients, >2 diseases for 96% of patients, and >3 diseases for 69% of patients. We found 17.8% of patients had a single G-probability >50% corresponding to a PPV of 41%. This compared to 45% of patients, and PPV of 64% reported in the original study. Accuracy of G-probabilities to discriminate clinician-defined disease was similar in our cohort (AUC of 0.86 95% CI 0.84-0.87) compared to the original study (AUC 0.84 95% CI 0.81-0.86). In 57% of patients in our cohort, the disease with the highest G-probability corresponded to the clinician-defined disease compared to 53% in the original study. Conclusion We were able to replicate several findings of the original study in a large independent cohort including calibration, high NPV, but PPV was lower, suggesting that G-PROB is most valuable as a tool to rule out diagnoses. Disclosure R.M. Hum: None. S.D. Sharma: None. M. Stadler: None. N. Nair: None. S. Viatte: None. C. Yap: None. J.H. Humphreys: None. A. MacGregor: None. M. Yates: None. M. Soomro: None. S.M. Verstappen: None. P. Ho: None. A. Barton: None. J. Bowes: None.