Large-Scale Population Screening for Type 2 Diadetes Based on TabNet and BRFSS 2021 Data

Type 2 diabetes is a global health crisis. Early screening is of great significance for disease prevention and control. Based on the data from the CDC BRFSS 2021 large-scale real population survey (236378 cases, 21 characteristics), this study systematically evaluated the performance of TabNet, XGBoost and random forest in diabetes screening. 50% cross validation results showed that the recall rate of TabNet reached 77.81% ± 2.02%, which was significantly better than XGBoost (35.92% ± 0.51%) and random forest (32.64%±0.39%), and the AUC reached 0.8043±0.0031. The age subgroup analysis showed that among the high-risk population aged 30-50, the TabNet recall rate was as high as 80.88%, 51.93% higher than XGBoost. SHAP interpretability analysis showed that self-rated health, hypertension and age were the core characteristics for predicting diabetes. This result shows that TabNet's attention mechanism can effectively capture complex feature interactions and significantly improve the recognition ability of patients on unbalanced data. This study shows that TabNet has significant advantages of high recall rate in large-scale real population diabetes screening, and has important clinical application value.