(1067) Identifying Environmental, Sociocultural, Behavioral and Biological Predictors for Diabetes Self-management Education Engagement Using Machine Learning Approaches

Objectives: Diabetes is a chronic disease characterized by insulin resistance that affects an estimated 37.3 million people in the US. Because diabetes symptoms are often overlooked, including weight fluctuation, abnormal thirst, abnormal urination, and fatigue, nearly 23% of adults with diabetes are undiagnosed. The Centers for Disease Control and Prevention (CDC) has recognized the importance of a tailored education plan for patients with or at risk of diabetes and offers Diabetes Self-Management Education (DSME) as a cost-effective solution. Despite the efforts from the US government and non-profit organizations, along with substantial evidence supporting the cost-effectiveness and clinical benefits of DSME on patients with diabetes, the utilization of DSME remains very limited. Hence, identifying the facilitators and barriers to receiving DSME among patients with diabetes is urgently needed. Therefore, we aim to investigate comprehensive predictors for DSME using machine learning (ML) models and develop algorithms for the engagement of DSME among Medicare beneficiaries with diabetes in environmental, sociocultural, behavioral, and biological domains based on the National Institute of Aging (NIA) Health Disparities Research Framework.

Methods: The Medicare Current Beneficiary Survey (MCBS) data from 2017 to 2019 was employed in this study. The presence of diabetes was assessed through self-reports. Medicare beneficiaries who reported having diabetes were then asked whether they had participated in a diabetes self-management course. A total of 95 variables were used in our study, including 11 variables of the environmental domain, 10 variables of the sociocultural domain, 58 variables of the behavioral domain, and 16 variables of the biological domain. In this study, five ML models were utilized, including Support Vector Machines (SVM), Random Forest (RF), Logistic Regression (LR), Gradient-boosted trees (XGB), and LASSO (least absolute shrinkage and selection operator) regression. To assess the performance of ML models, various evaluation metrics models were employed, including receiver operating characteristic (ROC), area under the curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F1 score. This study also adopted SMOTE (synthetic minority over-sampling technique) to deal with unbalanced class data. Recursive Feature Elimination (RFE) was employed as a method for selecting key variables.

Results: In our study, 49.64% of participants received DSME after exclusion. Among the five models, RF and XGB had better performance. The AUC for RF and XGB were 96% and 93% respectively. RF achieved an accuracy of 87%, sensitivity of 82%, specificity of 91%, PPV of 90%, NPV of 83%, and an F1 score of 86%. For XGB, the accuracy, sensitivity, specificity, PPV, NPV, and F1 score were 85%, 78%, 93%, 92%, 80%, and 84%, respectively. 37 key variables were selected according to the result of RFE, including 9 variables of the environmental domain, 2 variables of the sociocultural domain, 12 variables of the behavioral domain, and 14 variables of the biological domain. After excluding individuals with missing values, the results confirmed the robustness of our study.

Conclusions: This study identified comprehensive and critical factors associated with the engagement of DSME from environmental, sociocultural, behavioral, and biological domains using different ML models. By aligning these findings with the DSME guidelines, implementation can be personalized to diverse patient needs, integrate effective behavioral techniques, foster social support networks, enhance health literacy, and facilitate ongoing adaptation. This study found important facilitators and barriers to implementing DSME, suggesting that pharmacists may play an important role in identifying patients with these characteristics in advance for early and effective intervention.