Predictors of Effectiveness of First-line Biologic Disease Modifying Antirheumatic Drug(s) (bDMARD) or Targeted Synthetic DMARD (tsDMARD) Treatment with a Tumor Necrosis Factor Inhibitor in Patients with Rheumatoid Arthritis

Background: Given treatment guidelines and options, finding effective treatments to manage rheumatoid arthritis (RA) can be a challenge.

Objectives: To evaluate the application of machine learning (ML) to real-world data (RWD) to identify predictors of response to tumor necrosis factor inhibitor (TNFi) treatment among patients with RA previously naïve to biologic disease modifying antirheumatic drug(s) (bDMARD) or targeted synthetic DMARD (tsDMARD) therapy.

Methods: We used Optum® MarketClarity® data from Jan 2011-Mar 2019, which included integrated electronic health records (EHR) and claims data for >15M lives, to identify individuals diagnosed with RA receiving 1st line TNFi treatment. The primary outcome of ineffective TNFi response was defined using Curtis et al's algorithm (a validated claims-based algorithm for determining the effectiveness of TNFi medications in RA patients). Treatment was deemed to be ineffective if patients met any of the following criteria during one year of follow-up: 1. Inadequate adherence to index drug; 2. bDMARD switch or addition; 3. addition of new non-biologic DMARD; 4. increase in bDMARD dose or frequency; 5. >1 glucocorticoid joint injection; or 6. increase in dose of oral glucocorticoid. We examined 772 potential predictors drawn from claims data during the year prior to TNFi initiation. A Bayesian methodology was employed with an ensemble approach that examines a network of at least 128 predictive models.

Results: 2,018 patients met our eligibility criteria (74% female, mean age 54.6 years). TNFi treatment was determined to be ineffective in 71% of patients. The main predictor of TNFi ineffectiveness was ≥1 ER visit (odds ratio [OR] 0.79; 95% confidence interval [CI] 0.77,0.84]). Because ER visits was such a strong predictor, when ≥1 ER visit was removed from the model, abdominal pain became the strongest predictor of TNFi ineffectiveness (OR 0.51; 0.51,0.67). The strongest predictor of TNFi treatment effectiveness was prior use of methotrexate (OR 1.55; 1.55,2.04). The AUC median was 0.56 with maximum of 0.60.

Conclusions: ML methods can be leveraged to address questions in pharmacoepidemiology using RWD. Rather than a single predictive model, this ensemble modeling approach enables better sampling of the modelling space (including interactions among variables) and increases predictive power. It amplifies relationships best supported by the data and downplays spurious ones. However, even with applying ML methods to a rich EHR/claims integrated RWD source, it proved difficult to predict when 1st line treatment with a TNFi would be ineffective for patients with RA. Nonetheless, prior use of methotrexate increased the likelihood of TNFi effectiveness, while ≥1 ER visit and abdominal pain reduced the likelihood of effectiveness.