An XGBoost-based model for detecting undiagnosed type 2 diabetes using routine physical and lifestyle data from a multi-center Chinese population

A research team developed and validated an XGBoost machine learning model to identify individuals with undiagnosed type 2 diabetes using data from routine health checkups. The study included data from 11,382 individuals from 12 tertiary hospitals in China as the training set and 1,026 individuals as an independent test set. The final model used 12 predictors, with fasting blood glucose being the most influential (50.6%), followed by creatinine (6.6%), triglycerides (5.6%), age (5.1%), and LDL cholesterol (5.0%). The model achieved an AUC value of 77.2% (95% CI: 70.3%–84.1%) on the test set. Researchers concluded that the model demonstrates moderate predictive performance and has potential to be integrated into clinical practice as an auxiliary screening tool.