Objective: Comparison of the performance of two artificial intelligence models using the Decision Tree algorithm for predicting HbA1c ≥ 6.5% based on biochemical test results at Bac Kan General Hospital in 2026.
Methods: A cross-sectional study was conducted at Bac Kan General Hospital in 2026, involving 880 patients who underwent HbA1c and biochemical testing. Two Decision Tree models were developed using SPSS software based on the CART method, with a 70/30 training-to-testing data split. Model A utilized the variables of age, glucose, total cholesterol, triglycerides, AST, and ALT; Model B utilized age, total cholesterol, triglycerides, AST, and ALT (excluding glucose). The performance of both models was evaluated based on accuracy, sensitivity, specificity, and risk estimate.
Results: Among the 880 patients studied, the group with HbA1c ≥ 6.5% exhibited statistically significantly higher glucose and triglyceride levels compared to the group with HbA1c < 6.5% (p < 0.05). Model A demonstrated superior performance to Model B on the test set, achieving an accuracy of 79.8% versus 57.5%; sensitivities of 63.0% and 12.7%, respectively; specificities of 89.7% and 90.6%; and error estimates of 0.202 and 0.425. In Model A, glucose was the most important variable and served as the initial splitting node of the decision tree; when glucose was removed, triglycerides became the model's most important variable in Model B.
Conclusions: The Decision Tree algorithm can assist in predicting HbA1c levels ≥ 6.5% based on biochemical test results. Incorporating the glucose variable significantly improves the model's predictive performance, underscoring the crucial role of glucose in developing artificial intelligence models to screen for patients at risk of elevated HbA1c.