This study demonstrated excellent predictive capabilities for DSME engagement through the utilization of ML models. Among the ML models evaluated, our results indicated that RF has the best performance for DSME prediction, with an AUC of 94% and an accuracy of 85%. Additionally, we identified 74 of 95 variables as key variables, and racial/ethnic disparities in predictors for DSME engagement were found.

According to previous studies, certain individuals are less inclined to participate in DSME services, including males, non-White individuals, those with lower education levels, individuals of lower socioeconomic status, and those with more severe disease conditions.22 23 This is consistent with the findings of our studies. We found that income, education level, race, sex, and comorbidities (eg, cancer, stroke, depression, and chronic obstructive pulmonary disease) can serve as key variables associated with DSME participation.

Among the top 10 important predictors, half of them belong to behavioral domains. As recommended by organizations like the CDC,24 individuals with diabetes can effectively enhance their condition and overall health by adhering to evidence-based self-care and clinical preventive measures. These measures encompass actions such as annual foot and eye examinations, adopting a physically active lifestyle, refraining from smoking, and consistently monitoring blood sugar levels. Our studies effectively pinpointed these factors as pivotal indicators of DSME engagement, including knowing about managing diabetes, testing own blood for sugar/glucose, checking for sores on feet, foot examination, and smoking status. Previous studies have also highlighted the association between diabetes preventive care practices and DSME.25 26 A study using data from the Behavioral Risk Factor Surveillance System found that individuals who received DSME exhibited a greater likelihood of abstaining from smoking, conducting daily glucose testing, performing regular foot checks, and participating in leisure-time physical activities.26

In addition to the top 10 important predictors, our research revealed that limited mobility (ie, having difficulty stooping/crouching/kneeling, walking ¼ mile, writing/handling objects, lifting/carrying 10 pounds, extending arms above the shoulder) could serve as key predictors of DSME engagement, which is a new finding of this study. Older individuals with type 2 diabetes frequently experience skeletal muscle impairment, resulting in diminished muscle strength and compromised physical function.27 Diabetes-related limited mobility could further impede patients from participating in DSME activities. Moreover, having difficulty with instrumental activities of daily living (IADL), especially having difficulty using telephone and managing money, was identified as a novel predictor. IADL are activities that enable an individual to live independently in a community, and assessing IADL is useful for determining an individual’s cognitive function.28 Patients’ difficulty with IADL might impact their understanding of DSME activities, potentially reducing their willingness to participate. In summary, aside from barriers mentioned in the Consensus Report on DSME5 —including common health system or programmatic barriers, referring healthcare providers’ barriers, participant-related barriers, and environment-related barriers—patients' mobility abilities and difficulties in IADL should be specifically focused on and addressed.

Race is the 11th key predictor based on importance, and racial/ethnic disparities in predictors for DSME engagement were identified in this study. Focusing on the top 10 important variables, smoking status is the distinguishing variable for non-Hispanic Whites. The relationship between smoking and DSME was demonstrated previously,26 considering the relatively high prevalence of smoking among non-Hispanic Whites, more efforts should be taken to improve DSME engagement among White patients. For non-Hispanic Blacks, health status (compared with 1 year ago) and ever having any problems with kidneys were identified as unique variables. Poor health status is related to less DSME participation,11 thus extra measures should be taken to improve health status among Blacks given that racial/ethnic minority groups experience higher rates of illness across various health conditions compared with their White counterparts throughout the USA. To enhance DSME participation among Hispanics, additional efforts should focus on examining any problems with eyes due to diabetes, improving Medicaid coverage, and controlling BMI. For other racial/ethnic groups, addressing depression and sleep problems could be potentially beneficial in improving DSME engagement.

In contrast to other studies investigating risk factors for DSME engagement, our research held an advantage by encompassing an extensive range of about 100 variables across diverse domains. Even though DSME has been proven to offer substantial value and effectiveness, its potential success is hindered by the challenge of low utilization due to various barriers. Focusing more on person-centered care is one of the major guiding principles for the 2022 revision of the National Standards.8 Through the identification and resolution of these barriers, diabetes care and education specialists can address additional hurdles individuals might face and the benefits can be fully realized, which is consistent with the goal of National Standards by ADA. In addition, this study is the first study developing ML models based on the NIA Health Disparities Research Framework for DSME engagement. Studies based on the NIA Health Disparities Research Framework can effectively address health disparities by emphasizing key factors related to aging, providing a structured approach to monitor progress, enhancing the understanding of causal pathways, and expanding intervention targets to address modifiable factors.13 This comprehensive approach supports targeted efforts to reduce health disparities in the aging population.

There were several limitations in this study. First, approximately one-third of the variables in our analysis exhibited missing rates exceeding 10%, potentially constraining the capacity to establish accurate associations. However, we conducted a sensitivity analysis that involved excluding individuals with missing values for any variables. This evaluation confirmed the robustness of our results despite the missing data issue. Moreover, it’s essential to acknowledge that this research was carried out in the form of a cross-sectional study. Nonetheless, the selection of solely those patients already diagnosed with diabetes for further assessment of DSME involvement ensures that the sequence involves a diabetes diagnosis preceding DSME participation. This arrangement suggests that the necessity for temporality in drawing causal conclusions has been largely addressed. As a result, this distinctive research design indeed permitted the identification of potential causal inferences for projecting DSME engagement.