Machine Learning Assessment of Cost-Related Medication Non-adherence Among Canadians with Arthritis

Objectives Cost is a well-established barrier to treatment adherence, contributing to poorer health outcomes.[1] While cost-related nonadherence (CRNA) has been studied in the general population, it remains under-characterized among Canadians with arthritis, a group with distinct treatment needs and financial burdens. This study applied machine learning to identify and rank factors associated with higher probability of CRNA in this population and to estimate its prevalence and burden. Methods Data were sourced from the 2016 Canadian Community Health Survey. Responses from 23,687 individuals with arthritis were analyzed for CRNA, defined as self-reported skipping or not filling prescriptions due to cost in the past 12 months. Sixteen demographic and health/health-system variables were examined using case-weighted random forest models to estimate variable importance (via permutation-based assessment) in the overall and sex-stratified populations. Missing data were imputed using multiple imputations by chained equations. For enhanced interpretability and burden assessment, case-weighted multivariable logistic regression models (overall and sex-stratified) estimated the direction and magnitude of associations. Prevalence estimates and 95% confidence intervals (CIs) were calculated using survey weights and 1,000 bootstrap replicates. Results The weighted prevalence of CRNA was 7.5% (95% CI 6.9-8.1), with higher prevalence among females (8.5%, 95% CI 7.7-9.4) than males (6.1%, 95% CI 5.3-6.9). In random forest models, the most influential factors associated with higher probability of CRNA in the overall population were age, insurance coverage, household income, and satisfaction with life (Figure 1). Within the top quartile of factors for both males and females independently were age and insurance coverage. However, remaining top variables differed: among females, household income and education ranked highly, whereas among males, perceived health and satisfaction with life were more prominent (Figure 1). In the multivariable logistic regression, higher odds of CRNA were observed among females (OR=1.37, 95% CI 1.13-1.66), younger adults ≤34 years (OR=5.22, 95% CI 3.53-7.70), non-white individuals (OR=1.53, 95% CI 1.01-2.32), those with lower income (