The ensemble learning model is not better than the Asian modified CKD-EPI equation for glomerular filtration rate estimation in Chinese CKD patients in the external validation study

Objective To assess the clinical practicability of the ensemble learning model established by Liu et al. in estimating glomerular filtration rate (GFR) and validate whether it is a better model than the Asian modified Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) equation in a cohort of Chinese chronic kidney disease (CKD) patients in an external validation study. Methods According to the ensemble learning model and the Asian modified CKD-EPI equation, we calculated estimated GFRensemble and GFRCKD-EPI, separately. Diagnostic performance of the two models was assessed and compared by correlation coefficient, regression equation, Bland–Altman analysis, bias, precision and P30 under the premise of 99mTc-diethylenetriaminepentaacetic acid (99mTc-DTPA) dual plasma sample clearance method as reference method for GFR measurement (mGFR). Results A total of 158 Chinese CKD patients were included in our external validation study. The GFRensemble was highly related with mGFR, with the correlation coefficient of 0.94. However, regression equation of GFRensemble = 0.66*mGFR + 23.05, the regression coefficient was far away from one, and the intercept was wide. Compared with the Asian modified CKD-EPI equation, the diagnostic performance of the ensemble learning model also demonstrated a wider 95% limit of agreement in Bland-Altman analysis (52.6 vs 42.4 ml/min/1.73 m2), a poorer bias (8.0 vs 1.0 ml/min/1.73 m2, P = 0.02), an inferior precision (18.4 vs 12.7 ml/min/1.73 m2, P < 0.001) and a lower P30 (58.9% vs 74.1%, P < 0.001). Conclusions Our study showed that the ensemble learning model cannot replace the Asian modified CKD-EPI equation for the first choice for GFR estimation in overall Chinese CKD patients.