Metabolic Characteristics and Factors Associated with Prediabetes in Chinese Adults Based on Real-World Health Examination Data: A Cross-Sectional Study

Why It Matters

The findings pinpoint readily measurable factors—especially BMI and liver health—that can flag prediabetes early, enabling preventive actions and reducing future diabetes and cardiovascular costs. Machine‑learning validation demonstrates a scalable path for population‑level screening in primary care.

Key Takeaways

• BMI and fatty liver strongest predictors of prediabetes
• Lipid abnormalities (high TG, TC, LDL-C) linked to prediabetes
• Age and hypertension increase prediabetes risk
• Lower HDL-C and bilirubin inversely associated with prediabetes
• XGBoost model achieved AUC 0.94 for early detection

Pulse Analysis

Prediabetes, an intermediate stage of dysglycemia, affects a growing share of adults worldwide and signals heightened risk for type 2 diabetes and cardiovascular disease. In China, routine health examinations provide a rich, real‑world data source to map the metabolic landscape of this condition. A recent cross‑sectional analysis of 20,271 Chinese adults, collected between 2018 and 2024, compared individuals with normoglycemia to those meeting prediabetes criteria, uncovering a suite of clinical and biochemical markers that differentiate the two groups.

The study confirmed that prediabetic participants were older, carried higher body‑mass index (average 26.8 kg/m² versus 22.8 kg/m²), and were more likely to have hypertension and fatty liver disease. Lipid panels showed elevated total cholesterol, triglycerides, and LDL‑C, alongside reduced HDL‑C and total bilirubin. Multivariable logistic regression identified BMI (adjusted odds ratio 2.52), fatty liver (aOR 2.80), and dyslipidemia as independent risk factors, while higher HDL‑C and bilirubin were protective. Adding BMI to the predictive model lifted the AUC from 0.76 to 0.83.

Beyond traditional statistics, the researchers applied an XGBoost algorithm with SHAP interpretation, achieving an impressive AUC of 0.94. The model highlighted BMI, age, triglycerides, fatty liver, and total cholesterol as the most informative variables, with a nonlinear BMI effect that turned positive around 24 kg/m². These insights suggest that integrating simple anthropometric measures and routine lab tests into machine‑learning pipelines could enable earlier identification of high‑risk individuals in primary‑care settings, guiding targeted lifestyle or pharmacologic interventions and easing the future diabetes burden.