Background/Purpose: Interpretable machine learning (ML) method can identify factors associated with biological or targeted synthetic disease modifying antirheumatic drugs (b/tsDMARDs) nonadherence and nonpersistence for rheumatoid arthritis (RA) based on big data. This study is to develop and validate several interpretable ML models in evaluating the nonadherence and the risk of nonpersistence of b/tsDMARDs among older adults with RA.

Methods: This retrospective, new-user study used 5% national Medicare administrative claims data between 2012-2020, and included older (≥65) patients initiating b/tsDMARDs (the index date) between 2013-2019 and having a diagnosis of RA identified using the International Classification of Diseases, Ninth and Tenth revision codes. Nonadherence outcome was defined as medication possession ratio (MPR) < 80%, evaluated during the 12-month follow-up. Nonpersistence, defined as time from the index date until treatment gap ≥ 60 days observed during the follow-up, or censored for loss to follow-up. Patient-level data was split into 75% for model training and 25% for evaluation, with patients’ demographics and clinical characteristics measured in the baseline period. ML classification models (random forest [RF], eXtreme Gradient Boosting [XGBoost] and prediction rule ensembles [PRE]) evaluated the binary nonadherence outcomes and the ML survival model (random survival forest [RSF] and regularized Cox regression [RegCox]) assessed time on nonpersistence. Using the testing cohort, C-index and integrated Brier Score were reported for the ML survival model and AUROC was reported for ML classification models. Top influential predictive features for nonadherence and the risk of nonpersistence were identified according to variable importance (VIMP) procedure from both types of ML algorithms, respectively.

Results: We identified 3,927 eligible patients with RA (mean age 73 ±6; female 75%). Approximately 18.5% of RA patients had the risk of nonpersistence after initiating their index b/tsDMARDs (mean time to nonpersistence 1,110 ±670 days). RSF (C-index 0.6629, iBS 0.1629) and Reg-Cox model (C-index 0.6810) found index year, age, Elixhauser score, frailty score, inpatient visit, outpatient visits, the type of b/tsDMARDs, and CIRAS as top 8 influential predictors for nonpersistence. 53.65% of RA patients were adherent (MPR 0.72 [0.31]; 0.01-1) in the 12-month follow-up. Compared to ML-based classification models (PRE 0.5407 [0.509-0.5722]; RF 0.6029 [0.5715-0.6336]; XGBoost 0.6018 [0.5704-0.6326]) had comparable performance in predicting nonadherence. Frailty score, CIRAS, age, index_year and Elixhauser comorbidity index were ranked commonly as the top influential features for nonadherence.

Conclusion: We effectively developed ML-based survival and classification models to explain the nonadherence and the risk of nonpersistence of b/tsDMARDs, respectively for older adults with RA. There is potential to leverage ML models for understanding patients’ behavior of utilizing b/tsDMARDs treatment with a goal to improve RA care.

« Back to ACR Convergence 2025

ACR Meeting Abstracts - https://acrabstracts.org/abstract/interpretable-ensemble-machine-learning-explaining-nonadherence-and-the-risk-of-nonpersistence-of-targeted-disease-modifying-antirheumatic-agents-in-older-adults-with-rheumatoid-arthritis/