Comparative analysis of diabetes-related factors using machine learning on the CDC health indicator dataset: metabolic, lifestyle, socioeconomic and demographic domains

Diabetes mellitus is a chronic disease that continues to pose significant global health pressure, affecting a considerable proportion of the world’s population, with an estimated 537 million people affected worldwide in 2021, with an expected increase of 46% in the number of diabetic patients by 2045 [1]. Diabetes is strongly associated with lifestyle behaviors, metabolic conditions, and socioeconomic status. This paper presents a comprehensive machine learning analysis of the CDC Health Indicator Dataset ( n = 253,680), sourced from the 2015 Behavioral Risk Factor Surveillance System (BRFSS). Following systematic data preprocessing with feature binning and multimodal exploratory data analysis, including interactive correlation matrices, radar charts, Sankey diagrams, animated feature distributions, and other graphical representations to examine data distributions, relationships, patterns, and feature importance. Four supervised classification models were trained and evaluated, including Logistic Regression, Decision Tree, Random Forest, and XGBoost. All four machine learning models achieved comparable overall accuracy ranging from 83.06% to 84.74%. As 84.3% of the participants belong from the non-diabetic class so the accuracy is driven almost entirely by the majority class due to which accuracy alone could be misleading. The Macro-averaged F1 (0.383–0.396) and balanced accuracy (0.384–0.387) shows the four models performing far closer together than accuracy alone suggests, and the per-class recall reveals a near-total inability to detect the pre-diabetic class (recall ≈ 0.00) regardless of the model choice. This result was confirmed under fivefold stratified cross-validation, in which Class 1 recall was exactly 0.0000 in every fold for three of the four models, and the results persisted across both the original BRFSS encoded data and the binned dataset. All four models were then retrained under balanced class weighting. Weighting raised the recall for the Class 1 (pre-Diabetic Class) recall from 0.0000 to between 0.293–0.310 for three of the four models, where the precision for the Class 1 remained between 0.021 and 0.031 and the overall accuracy decreased to approximately 58.26–73.08%. Cross Validated Class 1 PR-AUC ranged from 0.0243–0.0370 against a base rate of 0.018. These results indicate that class weighting can reduce majority-class bias and improve minority-class sensitivity, but it does not by itself provide reliable pre-diabetic discrimination from the available non-glycaemic survey predictors. Feature importance was found to depend on the representation used rather than being a fixed property of the data. Under the binned representation, general health status and high blood pressure ranked highest; under the original BRFSS coding, in which BMI is retained as a continuous variable, BMI was the strongest predictor by a wide margin. Predictor rankings are therefore reported separately for each representation. Income and education were associated with diabetes prevalence, but because the data are cross-sectional and self-reported, these associations are reported without causal interpretation.