Quantifying behavioral contributions to diabetes risk using an explainable multiclass machine learning framework with BRFSS 2024

Abstract This study developed an explainable multiclass framework to evaluate discrimination among non-diabetes, prediabetes, and diabetes using 2024 Behavioral Risk Factor Surveillance System data ( $$n=453{,}241$$ ). A dual-feature-set design compared behavioral/health-status predictors alone with an expanded set including demographic and anthropometric variables across four classifiers. Logistic Regression Model-2 showed the strongest class-balanced performance (macro-F1 = 0.426; balanced accuracy = 0.512), while Decision Tree Model-2 achieved the highest prediabetes recall (0.338). XGBoost Model-2 achieved the highest weighted F1-score (0.775) and macro ROC–AUC (0.765), and showed better probability calibration than Logistic Regression, but failed to identify prediabetes in the held-out test set. SHAP analysis of XGBoost Model-2 showed that demographic predictors accounted for 45.21% of total diabetes-class mean absolute attribution, behavioral/health-status predictors for 42.71%, and anthropometric predictors for 12.08%. Exploratory geographic SHAP aggregation showed variation in model-attributed signals, but these patterns should not be interpreted as population-level prevalence, comparative risk, or causal geographic effects. Overall, the findings show that behavioral/health-status information remained a substantial component of diabetes-class model attribution after demographic and anthropometric predictors were included, while also highlighting the importance of class-balanced and class-specific evaluation in multiclass diabetes prediction.