Abstract
Failure to initiate new prescriptions for methotrexate (MTX) and biologics/tofacitinib (B/T) is a significant barrier to effective rheumatoid arthritis (RA) management. This study confirms that primary nonadherence is common, affecting approximately 40% of patients. By integrating electronic health records (EHR) with insurance claims data, researchers developed predictive models capable of identifying high-risk patients at the point of care, offering a pathway for targeted early interventions to improve treatment initiation.
Introduction: The Challenge of Primary Nonadherence
In the management of Rheumatoid Arthritis (RA), early intervention with disease-modifying anti-rheumatic drugs (DMARDs) is critical to preventing irreversible joint damage. While clinical research often focuses on "secondary nonadherence"—patients discontinuing a medication they have already started—this study addresses the overlooked issue of primary nonadherence. This occurs when a physician issues a prescription that the patient never actually fills.
Closing the loop on primary nonadherence is technically challenging because it requires linking two disparate data sources: clinical intent (recorded in EHRs) and actual patient behavior (recorded in pharmacy claims). Understanding the factors that drive this gap is essential for clinicians to ensure that prescribed therapies are successfully initiated.
Methodology and Data Integration
The researchers conducted a retrospective cohort study using the Optum de-identified Integrated Claims-Clinical dataset. This dataset covers approximately 120,000 RA patients across more than 50 healthcare provider organizations in the United States between 2007 and 2015.
Cohort Selection
The study focused on two primary treatment groups: a Methotrexate (MTX) cohort of 763 patients and a Biologics/Tofacitinib (B/T) cohort of 434 patients. A 12-month "clean period" was established to ensure that the prescriptions analyzed were truly new treatment initiations rather than refills.
Predictive Modeling
To identify predictors of nonadherence, the team evaluated over 400 variables, including demographics, structured EHR data (e.g., BMI, vitals), and unstructured clinical notes processed via Natural Language Processing (NLP). They employed LASSO (Least Absolute Shrinkage and Selection Operator) logistic regression to select the most impactful variables and avoid model overfitting. The cohorts were split into training (75%) and test (25%) sets for validation.
Key Insights: Why Patients Don't Fill Prescriptions
The study revealed surprisingly high rates of primary nonadherence: 36.8% for methotrexate and 40.6% for biologics/tofacitinib. Several key factors emerged as strong predictors of this behavior:
- Insurance and Cost: For the more expensive B/T cohort, insurance type was a dominant factor. Patients with commercial insurance were significantly more likely to fill their prescriptions compared to those on Medicare, likely due to the high out-of-pocket costs associated with Medicare Part D specialty tiers.
- Frailty and Complexity: In the MTX cohort, urinary incontinence was a powerful predictor of nonadherence (Odds Ratio = 7.96). The authors interpret this not as a direct cause, but as a proxy for "frailty" or the burden of managing complex medication regimens in elderly patients.
- Socioeconomic vs. Clinical Factors: Interestingly, specific RA symptoms (like swelling or tenderness) and pain scores were less predictive than demographic and administrative data. This suggests that the barriers to starting treatment are often systemic or socioeconomic rather than purely clinical.
Results and Model Performance
The study successfully developed two high-performing predictive models:
- Methotrexate Model: Achieved an AUC of 0.86 in the test sample. Key predictors included advanced age (80+ years), specific geographic regions, and polypharmacy (high count of active drug ingredients).
- Biologics/Tofacitinib Model: Achieved an AUC of 0.78. Primary predictors included age, insurance type, and prior treatment history for inflammatory conditions.
Both models demonstrated good calibration, meaning the predicted risk levels closely aligned with actual observed nonadherence rates across the patient population.
Discussion and Clinical Implications
While the study has limitations—such as the need to infer "fills" based on timing rather than direct prescription IDs—it provides a powerful proof-of-concept for real-time risk assessment.
The authors propose that these predictive models could be integrated directly into EHR systems. When a rheumatologist prescribes a DMARD, the system could generate a real-time "risk score." For patients identified as high-risk, the clinic could trigger an immediate intervention, such as a consultation with a nurse or pharmacist to address financial barriers or concerns about side effects before the patient leaves the office. By targeting these "ghost prescriptions," healthcare providers can significantly improve the chances of successful treatment initiation and long-term joint health for RA patients.
Summary updated