Background In rheumatoid arthritis (RA) it is of major importance to distinguish non-responders to TNF-alpha inhibitor (TNFi) treatment before start to prevent a delay in effective treatment, potential side-effects and unnecessary healthcare costs. We investigated the ability of al large set of inflammatory proteins to predict (absence of) response to biological treatment.
Objectives To develop a protein score predictive for response to TNFi treatment in RA and investigate its added predictive value over clinical parameters alone.
Methods In consecutive RA patients eligible for TNFi treatment as included in the BiOCURA registry, serum was collected before start of treatment and analyzed on 57 inflammatory proteins using xMAP technology. EULAR response was determined after three months. A supervised cluster analysis method, partial least squares (PLS) was used to select the best combination of proteins and cross-validation to gain a reproducible protein score. Multiple imputation was used to account for missing data of baseline clinical parameters. Relevant clinical parameters for EULAR good response were selected by performing a univariate (p<0.2) and multivariable backward selection (p<0.1). The predictive ability of the final model with and without the protein score was assessed using the area under the receiving operater curve (AUC-ROC), negative predictive values (NPV) and the reclassification index (NRI).
Results Response was determined for 171 of the 192 cases starting treatment. On top of CRP and DAS28 at baseline, PLS revealed 9 important proteins: sCD14, IFNγ, MCP1, MIP1b, MIP3b, TARC, sTNFRI, sTNFRII and TSLP. These markers were able to explain 31.5% of the variance in DAS28 at 3 months.
Final models for prediction of TNFi response included baseline DAS28, naivety for bDMARDs, HAQ, RF positivity, concomitant MTX and GC use. The protein score did not improve the AUC-ROC of 0.80 (0.73-0.87), however, when the predefined cut-off for a NPV≥0.9 was set, the addition of the protein score resulted in the classification of 30 extra patients in the low probability category (table). An improved classification was observed of 24.2% and -4.8% for patients with and without a response respectively (NRI=19.42%).
Conclusions We showed that a combination of a protein score and clinical variables is able to predict absence of EULAR good response to TNF inhibiting treatment and can classify more patients at baseline in the appropriate risk category than clinical variable alone. This protein score may therefore contribute to a more patients tailored treatment, leading to a better usage of the available resources. In the near future these findings will be validated externally.
Acknowledgements The Society for Rheumatology Research Utrecht (SRU) consists of University Medical Center Utrecht, Antonius Hospital Nieuwegein and Utrecht, Diakonessen Hospital Utrecht, Meander Medical Center Amersfoort, Sint Maartenskliniek Woerden, Hospital St. Jansdal Harderwijk, Tergooi Hospital Hilversum, Flevo Hospital Almere.
Disclosure of Interest None declared