Background/Purpose: Methotrexate (MTX) is the preferred initial disease-modifying drug (DMARD) for rheumatoid arthritis (RA). However, up to 50% of patients respond inadequately to MTX. Clinically useful predictors that effectively identify patients with RA who are likely to respond to MTX are lacking. We aimed to identify clinical predictors of response to MTX among DMARD-naïve patients with RA using machine learning methods.

Methods: Using the resources of the Clinical Study Data Request Consortium (CSDR) we identified 14 randomized clinical trials (RCT) involving 3,003 patients with RA who were randomized to MTX monotherapy or placebo plus MTX. Data were accessed through the Vivli Center for Global Clinical Research Data. Studies with available Disease Activity Score with 28-joint count (DAS28) at baseline, 12 and 24 weeks were included, resulting in exclusion of one RCT. Patients on prior conventional or biologic DMARDs were excluded. Latent class mixed modeling of response to MTX was applied. Patient groups with similar trajectories were compared by pretreatment baseline characteristics: socio-demographics (age, sex, race), baseline DAS28 with erythrocyte sedimentation rate [ESR] and its individual parameters, C-reactive protein, 66 swollen joint count (SJC66), 68 tender joint count (TJC68), RA duration, baseline use of glucocorticoids, health assessment questionnaire [HAQ] score, and serologic status (positive for rheumatoid factor [RF] or anticitrullinated protein antibodies [ACPA]). Proximity imputation was used to impute missing values. Lasso analysis and random forests were used to identify predictors of MTX treatment response. RCT study indicators were included in the model to adjust for heterogeneity among studies.

Results: A total of 1,478 DMARD-naïve patients from 13 RCTs were included. Mean age 50.5 years, 79% female, 84% RF positive, 87% ACPA positive, median RA duration 1.9 years, mean baseline DAS28-ESR 6.7, mean baseline HAQ 1.6. We identified 3 trajectory groups of response to MTX based on mean change in DAS28-ESR between baseline and 24 weeks: Class 1, “responders”, >1.2 unit change per 3 months; Class 2 and 3 “non-responders” < 1.2 unit change per 3 months in each. Using Lasso methods, RA duration, baseline DAS28-ESR, ESR, CRP, SJC66, age, and patient global assessment of disease activity (PtGA) were important predictors of response to MTX. The discrimination of responders from non-responders was acceptable: area under the curve (AUC) 0.70. Concordantly, using Random forests, the top five predictors of response to MTX were baseline DAS28-ESR, RA duration, SJC66, TJC68 and PtGA (AUC 0.69). Sex, race, HAQ score, seropositivity and use of glucocorticoids were not predictive of response to treatment with MTX.

Conclusion: Several baseline clinical characteristics, including DAS28-ESR, RA duration, PtGA, SJC66 were predictive of response to MTX and should be considered during the decision-making when initiating MTX in DMARD-naïve patients with RA.

Acknowledgment: This abstract is based on research using data from data contributors UCB and Roche that has been made available through Vivli, Inc. Vivli is not in any way responsible for, the contents of this publication.

« Back to ACR Convergence 2021

ACR Meeting Abstracts - https://acrabstracts.org/abstract/clinical-predictors-of-response-to-methotrexate-in-patients-with-rheumatoid-arthritis-a-machine-learning-approach-using-clinical-trial-data/