Predicting Cost-Related Medication Adherence in Patients with Cognitive Impairment: A Machine Learning Approach

OBJECTIVES: Cognitively impaired patients struggle with medication adherence due to memory and comprehension issues, often worsened by financial constraints. Urgent innovation is needed to address this issue. This study aimed to use machine learning methods to predict cost-related medication nonadherence (CRMN) in such patients.

METHODS: A retrospective cross-sectional study was conducted using data from the National Health Interview Survey (NHIS), 2011–2022. CRMN is defined as failing to obtain prescribed medications due to financial constraints, measured by three NHIS prompts. Predictors included demographic, socioeconomic, and health-related factors. One-third of the study population was assigned to the training group, and two-thirds was assigned to the test group. Using the training group, six machine learning models were developed, including Logistic Regression, K-Nearest Neighbors (KNN), Random Forest (RF), Gradient Boosting Machine (GBM), XGBoost, and LightGBM. Model performance was assessed using area under curve (AUC), accuracy, sensitivity, and specificity on the test group. Additionally, SHapley Additive exPlanations (SHAP) analysis was used to identify the most important predictors.

RESULTS: 10,945 individuals were included in this study, with 3,648 assigned to the training group and 7,297 to the test group. Among the six models, GBM showed the best predictive performance, with the highest AUC of 0.749 and the highest accuracy of 0.818. Sensitivity and specificity varied across models, with XGBoost having the highest sensitivity of 0.511 and GBM having the highest specificity of 0.976. Using the top-performing model GBM, SHAP analysis showed that the five most important predictors were age (1.47), health status (0.61), income (0.60), education (0.49), and insurance status (0.29).

CONCLUSIONS: In this study, we developed a GBM model to predict CRMN in patients with cognitive impairment. Moreover, understanding the most important predictors such as age, health status, income, education, and insurance status, can guide targeted interventions and support strategies to improve medication adherence and overall patient outcomes.