POS0509 DEVELOPMENT OF A MULITNOMIAL PREDICTION MODEL OF TREATMENT RESPONSE TO ETANERCEPT IN A MULTI-CENTRE COHORT OF PATIENTS WITH ESTABLISHED RA

Abstract

BackgroundTreatment response in rheumatoid arthritis (RA) is assessed through EULAR response groups of good, moderate, and poor response. Clinical prediction models from the literature typically frame this as a binary model, to differentiate poor from good and moderate responders. Here, we develop a multinomial model, to predict each group separately, after 3 months on the anti-TNF drug Etanercept (ETN).ObjectivesDevelop and validate a multinomial prediction model of treatment response to ETN in RA, based on baseline clinical covariates.MethodsWe identified patients treated with ETN or biosimilars (N = 778) from the Biologics in RA Genetics and Genomics Study Syndicate (BRAGGSS). Response groups were derived from the CRP based 4C-DAS28 at baseline and 3 month follow up, yielding 310 good, 320 moderate, and 148 poor responders. A multinomial logistic regression model was fitted, using good responders as reference category. Multiple imputation by chained equations was used to impute missing data, and models were internally validated via bootstrapping. We report model accuracy, as well as calibration, and compare effect sizes across response groups. Table 1shows the baseline statistics, and odds ratios for the included covariates.Table 1.Baseline covariate statistics and odds ratios (in bold: significant at p < 0.05); HADS: Hospital Anxiety and Depression ScaleVariableMean (± SD)ORModerate [95% CI]pORPoor [95% CI]por % YesSwollen Joint8.84450.980.350.948e-3Count (SJC)(± 5.20)[0.95 1.02][0.89 0.98]Tender Joint14.68771.076e-61.050.01Count (TJC)(± 6.74)[1.04 1.10][1.01 1.08]General Health74.74291.000.60.981e-3Visual Analog Scale (GHVAS)(±17.79)[0.99 1.01][0.97 0.99]CRP19.07391.000.220.990.26(±25.07)[1.00 1.01][0.98 1.00]BMI30.30351.000.481.000.41(±23.28)[0.99 1.01][0.99 1.01]Age of47.33301.010.121.020.06onset(±13.86)[1.00 1.03][1.00 1.04]Disease9.94011.000.840.990.45duration(±10.35)[0.98 1.02][0.96 1.02]HAQ1.60851.480.022.951e-6(± 0.65)[1.06 2.08][1.91 4.54]HADS-Anxiety8.08681.040.191.060.12(± 4.54)[0.98 1.10][0.99 1.13]HADS-Depression7.38411.060.120.970.55(± 4.02)[0.99 1.13][0.89 1.06]Concurrent81.49%0.412e-40.520.03DMARD[0.26 0.66][0.28 0.94]Female78.66%1.390.121.110.71[0.92 2.10][0.65 1.87]Seropositive77.89%0.540.020.470.01[0.33 0.89][0.26 0.86]1st Biologic90.62%1.060.860.480.03[0.55 2.06][0.24 0.94]ResultsAdjusted for optimism, the multinomial model achieves an accuracy of 50.7% (IQR: 50 – 51.3%), with calibration slopes of 0.574 (IQR: 0.569 - 0.579) and 0.534 (IQR: 0.525 - 0.544) for moderate and poor response, respectively. Figure 1 shows a comparison of odds ratios (OR) for the different outcome groups. The Health Assessment Questionnaire (HAQ) score is the biggest driver of both moderate and poor response. Previous biologic treatment also predicts poor but not moderate response. Compared to the multinomial model, a binary model, that discriminates poor from moderate and good responders, underestimates the effect size of HAQ.Figure 1.Odds ratios of FIRSTBIO and HAQ for moderate and poor response. Size of crosses indicate 95% confidence intervals.ConclusionThe model predicts EULAR response groups moderately well but is poorly calibrated, which can partly be explained by the generally higher sample size requirement of multinomial modelling. In the multinomial model, moderate and poor response is largely driven by the same covariates, which leads to blurred boundaries between good and poor responders, when response groups are merged to create a binary problem. Future research should consider the most appropriate model choice to describe data, including the use of multinomial instead of binomial models. More research and bigger sample sizes are required to improve on this multinomial model.Disclosure of InterestsMichael Stadler: None declared, Stephanie Ling: None declared, Nisha Nair: None declared, John Isaacs Speakers bureau: Abbvie, Gilead, Roche, UCB, Grant/research support from: GSK, Janssen, Pfizer, Kimme Hyrich Speakers bureau: Abbvie, Grant/research support from: Pfizer and BMS, Ann Morgan Speakers bureau: Roche/ Chuga, Consultant of: GSK, Roche, Chugai, AstraZeneka, Regeneron, Sanofi, Vifor, Grant/research support from: Roche, Kiniksa Pharmaceuticals, Anthony G Wilson: None declared, Darren Plant: None declared, John Bowes: None declared, Anne Barton Grant/research support from: Pfizer, Galapagos, Scipher Medicine, and Bristol Myers Squibb.