Background: Cost-related medication non-adherence (CRN) is a persistent and serious challenge among the elderly population in the US. For elderly patients with limited economic means at increased risk of hospitalization, CRN may elevate the risk of repeated hospitalizations and emergency department visits, and lead patients into a downward spiral of worse heath and higher non-adherence. Literature suggests many potential risk factors for CRN; however, there is no comprehensive framework of the risk factors for CRN, in part because the relative importance of those risk factors is unknown. Machine-learning Random Forests (RF) is a highly flexible model which often provides high prediction accuracy; most importantly, the model provides a measure of which variables have the greatest influence on the outcome. We aim to develop a machine-learning prediction model of CRN by assessing a comprehensive set of risk factors for CRN using importance of variable (IOV) analysis with a RF model.
Methods: Older Medicare patients (age>=50) with at least 1 hospitalization in the past two years were identified using the 2014 Health and Retirement Study (HRS). HRS is a nationally representative sample of older adults with survey questionnaires on health, behavior, and socio-economic status. CRN was self-reported. A wide range of predictors, including socio-economic status; age; gender; race/ethnicity; educational attainment; ratio of Social Security (SS) income to the total income; Medicare-Medicaid dual eligibility; comorbidities including diabetes, heart disease, stroke, depression, and cancer; Activities of Daily Living (ADLs); and Instrumental Activities of Daily Living (IADLs) were extracted. An IOV analysis was conducted using the Mean Decrease in Accuracy (MDA) permutation approach with an RF model.
Results: Among the 3,781 older Medicare patients at increased risk of hospitalizations who were surveyed, 504 (13.3%) reported CRN. The IOV analysis shows that the top predictors for CRN by MDA values in descending order are: age (50), IADL (40), ADL (39), Medicare-Medicaid dual eligibility (20), ratio of SS income to total income (18), black race (14), depression (8), Hispanic ethnicity (8), diabetes (8), stroke (5), respectively.
Conclusions: Age, functional status, and a wide range of socio-economic variables along with comorbidities contribute to CRN critically and variably. These variables, when used jointly, can provide personalized predictions of CRN with ranking of importance in influencing CRN, which can be used by practitioners and policymakers to identify those patients at the higher risk of CRN.