Predicting Prescription Drug Adherence and Treatment Gaps for Medicare Beneficiaries with Chronic Conditions: A Comparison between Traditional Models and Machine Learning Algorithms
Over the past few decades, new pharmaceutical treatments including specialty drugs have offered new possibilities for patients with serious, chronic, or life-threatening diseases for whom prior treatments were ineffective, highly toxic, or previously unavailable. However, real-world studies have shown that patients are poorly adherent to these innovative drugs, particularly in the Medicare population. While medication adherence is already a nuanced, multifaceted phenomenon, Medicare patients often face unique circumstances (limited financial means, cognitive impairments, polypharmacy, etc.) that further complicate adherence. Despite this, little evidence exists on which Medicare beneficiaries are most susceptible to non-adherence or how best to identify them. Machine learning is one possible avenue for answering these questions. Machine learning holds the promise of having better predictive accuracy since it relies on a computer to learn aforementioned complex and non-linear interactions between variables of interest by minimizing errors between predicted and observed outcomes. Our first-of-its-kind study will use 100% Medicare Chronic Conditions Warehouse data to compare the predictive accuracy for medication non-adherence among Medicare beneficiaries with select chronic conditions using traditional logistic regressions and machine-learning algorithms. Identifying the model with the highest predictive accuracy can hopefully be used to target medication adherence interventions to certain patients, thereby maximizing the impact of limited health resources. Our findings from this study will be used to obtain federal funding for future research to identify patients at high risk for non-adherence and target interventions for these patients.