Machine Learning Models Identify Patients With Rheumatoid Arthritis At Risk for Treatment Nonadherence

Key Clinical Summary

• Interpretable machine learning models using Medicare claims modestly predicted rheumatoid arthritis (RA) medication nonadherence (area under the curve [AUC] ~0.63) and nonpersistence (C-index up to 0.67) among adults ≥65 years initiating advanced disease-modifying anti-rheumatic drugs (DMARDs).
• In this cohort (n = 3927), only 53.7% were adherent at 12 months; 18.5% experienced nonpersistence, highlighting substantial gaps in long-term treatment adherence.
• Predictive insights may support targeted interventions and proactive resource allocation to improve adherence and optimize RA outcomes.

Machine learning models using Medicare data can identify patients with RA at risk for medication nonadherence and nonpersistence, enabling targeted interventions and more efficient allocation of clinical resources to improve long-term treatment adherence, according to a study published in Clinical Therapeutics.

“In this present study, we leveraged several novel interpretable machine learning models, including machine learning classification models and machine learning based survival models to help us understand the patient’s medication taking behaviors associated with advanced targeted disease-modifying antirheumatic drugs for the a national Medicare population of older adults aged 65 years and above,” wrote the authors.

Investigators drew from a 5% sample of Medicare claims from 2012 to 2020 and, after applying inclusion and exclusion criteria, identified 3927 patients aged 65 years or older who newly initiated one of these agents between 2013 and 2019.

Most patients started a tumor necrosis factor inhibitor (TNFi) biologic DMARD (57.6%), followed by non-TNFi biologics (38.3%) and Janus kinase (JAK) inhibitors or other targeted synthetic DMARDs (4.1%). The cohort had a mean age of 73.4 years, was predominantly women (75.2%) and White (84.9%), and most commonly resided in the South (43.6%). Comorbidity burden was substantial, with mean frailty, claims-based RA severity, and Elixhauser scores of 0.19, 7.68, and 5.93, respectively. Osteoarthritis was present in 72.3%, back pain in 60.5%, osteoporosis in 34.1%, depression in 24.9%, interstitial lung disease in 20.0%, and serious infection in 11.2%. Oral glucocorticoids were used by 71.5%, opioids by 58.5%, and methotrexate by just over half of patients.

At 12 months, 53.7% of patients were adherent to therapy, defined as a medication possession ratio of at least 80%, with a mean medication possession ratio of 0.72. Predictive performance for nonadherence was modest across machine learning models. Random forest achieved an AUC of 0.6315, XGBoost 0.6277, and prediction rule ensembles 0.6271, with no significant differences between models. For nonpersistence, defined as a treatment gap of at least 60 days, 18.5% of patients experienced an event, while 81.5% remained persistent over a mean follow-up of 1008 days. Survival-based models performed somewhat better, with concordance indices of 0.661 for random survival forest and 0.670 for XGBoost survival, both outperforming the rule-based model at 0.634.

“Clinicians and health care personnel may benefit from these study findings to consider the development of targeted patient interventions, with a goal of efficiently allocating resources in advance to improve RA patients' medication taking behavior,” concluded the study authors.

Reference

Huang Y, Agarwal SK. Interpretable ensemble machine learning prediction of nonadherence and the risk of nonpersistence of targeted disease-modifying antirheumatic agents in older adults with rheumatoid arthritis. Clin Ther. doi:10.1016/j.clinthera.2026.01.00