The Comparison of Accuracy between Traditional Logistic Regression and Gradient Boosting Classifier Model of Artificial Intelligence in Predicting Diabetes Kidney Disease in Chinese – A 4-year Longitudinal Study

Abstract

Abstract Background: Type 2 diabetes (T2D) has been increasing recently in Taiwan which causes 43% of the total population of dialysis. In the present study, our goal was to compare the accuracy of logistic regression (LR) and gradient boosting classification (GBC) of artificial intelligence in predicting diabetes kidney disease (DKD) in a Chinese cohort. Methods: Totally, there were 365 men and 320 women with T2D enrolled and followed for four years. They were further divided into quintiles according to the estimated glomerular filtration rate (eGFR). Both LR and GBC were used to estimate the future DKD. Simple correlation was applied to evaluate the correlation between factors and eGFR at the end of follow-up (eGFR-FU). Sixty percent participants, as training group, were randomly sampled. The others were the validation group. The equations obtained from the training group of both methods are applied to calculate the receiver operation curve (ROC) of the validation group. Results: At the end of the follow-up, the eGFR-FU significantly different in both genders. The baseline eGFR is negatively related to age, duration of T2D, low density lipoprotein, ALT, systolic blood pressure, but positively related to fasting plasma glucose (FPG) and creatinine in men. In women, the relationship of FPG disappears. The ROC for LR is 0.88 and for GBC is 0.97 for men, and 0.82 and 0.94 for women, respectively. Both findings reach statistically significance. Conclusion: In conclusion, GBC could provide a better prediction compared to traditional LR in patients with T2D followed up for 4 years.