Background/Purpose:

As a step towards personalized medicine, we seek to identify patients with new-onset RA who are likely to remain well on MTX monotherapy. Patients unlikely to persistently respond could potentially avoid adverse effects such as pain, functional impairment, structural damage, reduced work ability, or co-morbidities if offered alternative treatments already at diagnosis.

This project aims at assessing the performance, and marginal gains, of different statistical approaches to modelling predictions of persistence to methotrexate DMARD monotherapy at 1 year after RA diagnosis in patients with new-onset RA. Here, we report first results.

Methods:

A cohort of incident RA diagnosed 2006-2014, starting treatment with MTX and DMARD monotherapy, and with clinical and treatment data available from diagnosis, was identified through the Swedish Rheumatology Quality (SRQ) register. Through linkages to nationwide population health and demographics registers, information on age, gender, educational level, income, hospital admissions and outpatient visits (coded using ICD10-codes), and prescribed drugs (coded using ATC codes) was collected. With regards to previous medical conditions and drug use, we compiled three sets of covariates, with increasing complexity: A) including 20 a priori defined co-morbid conditions only, B) including all ICD and ATC codes irrespective of time before RA, and C) including all ICD and ATC codes but with each ICD codes assessed in three different time periods before RA (<1 year, 1-4.9, and 5-10 years before RA). For B) and C), the ICD and ATC codes were further included at four levels of resolution (1 to 5-digit covariates). The outcome was defined as remaining on MTX as DMARD monotherapy, without any other type of DMARDS added or switched to, after 12 months.

We first assessed the association between all variables in covariate set C and the outcome in a univariate logistic regression. We then computed a logistic regression model for only gender and age as a baseline predictor, followed by L1-regularized logistic regression (“Lasso”) models based on the full covariate sets A, B, and C, respectively. Predictive capacity is estimated as average ROC AUC under 10-fold nested cross-validation

Results:

A total of 6225 patients with new-onset RA starting MTX as DMARD monotherapy were included. After 1 year, 4497 (72%) remained on MTX DMARD monotherapy. In the association analysis, 254 of the 1449 covariates had a p-value < 0.05. The logistic regression model with age and sex as the only covariates had an average ROC AUC of 0.596. The Lasso models showed mean ROC AUC values of 0.634 for set A, 0.650 for set B and 0.646 for set C.

Conclusion:

Prediction of persistence to MTX treatment is difficult and advanced analytical methods based on diagnostic codes and co-medication can potentially increase ROC AUC as compared to a baseline model. We are currently decomposing the main outcome into different subcomponents, e.g. early stopping due to side effects, to understand predictive potential more granularly and are exploring several machine learning methods such as random forest and deep learning to improve predictive performance.

« Back to 2018 ACR/ARHP Annual Meeting

ACR Meeting Abstracts - https://acrabstracts.org/abstract/predictors-of-persistence-to-methotrexate-treatment-in-ra-assessment-of-different-modelling-approaches/