Characteristics of participants
A total of 2449 participants were finally included in current analysis, their characteristics are described in Table 1. The overall median age was 57.00 years, and 61.90% were male. The mean mGFR was (70.30 ± 27.91) mL/min per 1.73 m2, and 12.05% participants were diagnosed with CKD. The prevalence of diabetes and hypertension was 19.31% and 44.22%, respectively. After randomly grouping, 1712 and 737 participants were assigned to development and validation datasets. Only the sex proportion of the two datasets was significantly different. The mean age, Scr, CysC, and prevalence of CKD, diabetes, and other diseases were all similar in the two datasets.
Table1. Characteristics of study participants, overall and by dataset
Characteristic
|
Overall (n=2449)
|
Development sample (n=1712)
|
Internal validation sample (n=737)
|
P value*
|
Age (years)
|
57.00(21.00)
|
57.00(22.00)
|
56.00(21.00)
|
0.49
|
Male (n, %)
|
1516 (61.90)
|
1037(60.57)
|
479(64.99)
|
0.04
|
Height (cm)
|
165.00 (12.00)
|
165.00 (12.00)
|
165.00 (12.00)
|
0.27
|
Weight (kg)
|
64.00 (17.50)
|
64.00 (17.00)
|
64.00 (17.00)
|
0.80
|
BSA (m2)
|
1.70±0.18
|
1.70±0.18
|
1.70±0.18
|
0.87
|
BMI (kg/m2)
|
23.63(4.55)
|
23.63(4.64)
|
23.56(4.61)
|
0.38
|
mGFR (mL/min per 1.73m2)
|
70.30±27.91
|
70.54±27.62
|
69.75±28.57
|
0.52
|
Scr (m mol/l)
|
81.60(52.40)
|
81.25(52.15)
|
83.00 (55.20)
|
0.40
|
Scys (mg/dl)
|
1.40(38.00)
|
1.45(38.00)
|
1.34(37.49)
|
0.47
|
BUN(mg/dl)
|
5.80(3.21)
|
5.77(3.22)
|
5.84(3.21)
|
0.41
|
ALB (mmol/l)
|
39.10(41.08)
|
38.90(41.13)
|
39.30(40.78)
|
0.17
|
Diabetes a (n, %)
|
473 (19.31)
|
319(18.63)
|
154(20.90)
|
0.20
|
Hypertension b (n, %)
|
1083(44.22)
|
749(43.75)
|
334(45.32)
|
0.47
|
CKD (n, %)
|
295(12.05)
|
198(11.57)
|
97(13.16)
|
0.27
|
CRI (n, %)
|
13(0.53)
|
9(0.53)
|
4(0.54)
|
0.96
|
DN (n, %)
|
24(0.98)
|
18(1.05)
|
6(0.81)
|
0.58
|
NS (n, %)
|
18(0.73)
|
13(0.76)
|
5(0.68)
|
0.83
|
GN (n, %)
|
89(3.63)
|
63(3.68)
|
26(3.53)
|
0.85
|
Abbreviations: BSA, body surface area; BMI, body mass index; mGFR, measured glomerular filtration rate; Scr, serum creatinine; Scys, serum cystatin C; BUN, blood urea nitrogen; ALB, plasma albumin; CKD, chronic kidney disease; CRI, chronic renal insufficiency; DN, diabetic nephropathy; NS, nephrotic syndrome; GN, glomerulonephritis.
Note: Data are shown as mean ± SD for continuous variables with normal distribution, and mean (IQR) for skewed-distribution variables, numbers and percentage was shown for categorical variables.
* t test was used for normal-distribution continuous variables, and Wilcoxon Rank Sum Test used for skewed-distribution continuous variables. For categorical variables Chi-squared test was used for between-group comparison.
Newly developed equations
A multivariable analysis modeling with least-squares regression was performed to screen in the log-mGFR-related variables, variables were on natural logarithms scale or original status. Last, age, sex, log-Scr, CysC, log-BUN, log-weight were better correlated with mGFR, and they were included in a multivariable analysis modeling with log-mGFR. The backward method was used to filter variables, and adjusted R2 and AIC were calculated to evaluate the performance of the model. With the filter of candidate variables, adjusted R2 of model increased. Model 3 with 4 variables, age, sex, log-Scr and CysC, had a similar adjusted R2 to Model 4 which removed CysC (0.669 versus 0.672) (Supplementary Table S1). Therefore, Model 3 and Model 4 were selected for eGFR equation development. We developed 9 eGFR equations (Supplementary Table S2), and compared them in the two datasets (further details are available in Supplementary Table S3). After evaluating their performance, 3 equations were deemed as the best-performing equations for general use (Table 2). Two of them were newly developed, named “SC-GFRcr-cys” and “SC-GFRcr”. One was derived from CKD-EPI4race equation, named “SC-CKD-EPIcr”.
Table 2 The new equations for estimating GFR
Name
|
Equations
|
SC-GFRcr
|
1868.266×Scr (-0.642) × (0.993) Age×0.908 (if female)
|
SC-CKD-EPIcr
|
if Scr ≤ 0.67(for female): 128×(Scr/0.67)-0.364×0.992Age
|
if Scr>0.67(for female): 128×(Scr/0.67)-0.677×0.992Age
|
if Scr ≤ 0.92(for male): 115×(Scr/0.92)-0.364×0.992Age
|
if Scr>0.92(for male): 115×(Scr/0.92)-0.677×0.992Age
|
SC-GFRcr-cys
|
1811.756×Scr(-0.643)×(1.002)Cys×(0.992)Age×0.908(if female)
|
Abbreviations: Scr, serum creatinine; Scys, serum cystatin C; SC-GFRcr, Southeast Chinese modification
of eGFR equation based on serum creatinine; SC-CKD-EPIcr, Southeast Chinese modification of Chronic
Kidney Disease Epidemiology based on serum creatinine; SC-GFRcr-cys, Southeast Chinese modification
of eGFR equation based on serum creatinine and serum cystatin C.
Performance of the previously reported equations and new developed equations
A total of 16 existing eGFR equations developed among Asian populations were listed in Supplemental Table S4, and their performance was summarized in Supplemental Table S5. There were 11 equations based on Scr, and 5 equations introduced CysC. The P30 of CKD-EPIco, CKD-EPIcd2, MDRDmac, and Xiangya were 74.72%, 75.46%, 73.74%, and 74.32%, meanwhile, their Bias were 2.95, 0.52, -1.24, and 1.39, respectively. While, the P30 of the other 12 equations were all < 70%, and their Bias were all > 5 ml/min/1.73m2. Thus, we included the CKD-EPIco, CKD-EPIcd2, MDRDmac and Xiangya equations in comparison with newly developed equations.
The P30 of three new eGFR equations were all ≥ 75% in both development and internal validation datasets, which is considered sufficient for Good clinical decision-making according to the 2002 K/DOQI benchmark (Table 3)(9). In both two datasets, the SC-GFRcr equation had lower Bias (-0.32, 0.26 mL/min/1.73 m2), and lower IQR (20.55, 20.83 mL/min/1.73 m2) than the other two equations. The SC-CKD-EPIcr equation had highest P30 (79.44%, 78.45%), and second lowest RMSE (17.18, 17.79). As for the SC-GFRcr-cys equation, the performance was as following: Bias (-0.43, -0.01 mL/min/1.73 m2), IQR (20.78, 20.25 mL/min/1.73 m2), P30 (78.33%, 76.12%), RMSE (17.36, 17.76), respectively. All three new equations had slightly higher P30 than previously reported equations. The Bias of two new equations, SC-GFRcr and SC-GFRcr-cys, and 1 reported equation, CKD-EPIcd2, were all lower than 1.0 mL/min/1.73 m2, and that of other equations were higher. The IQR of all new equations ranged from 20.00 to 21.00, which was similar to that of CKD-EPIco, but higher than those of other 5 reported equations. Bland-Altman plots of all equations are showed in Figure 1-7, and the LOA were (0.14 ± 34.63), (3.56 ± 34.02) and (0.30 ± 34.41) mL/min/1.73m2 for SC-GFRcr, SC-CKD-EPIcr and SC-GFRcr-cys equations separately. The SC-CKD-EPIcr equation had lowest LOA compared with other equations.
Table 3. Performance of GFR estimation equations
Equations
|
Development dataset (N=1712) mGFR *= 70.54 mL/min/1.73 m2
|
Internal validation dataset (N=737)
mGFR* = 69.75 mL/min/1.73 m2
|
|
|
eGFR
|
Bias †
|
IQR ‡
|
P30§
|
RMSE
|
eGFR
|
Bias †
|
IQR ‡
|
P30§
|
RMSE
|
|
Previous reported equations
|
|
|
|
|
|
|
|
|
|
CKD-EPIco
|
73.41
|
2.87
|
20.35
|
75.47
|
17.22
|
72.88
|
3.13
|
20.47
|
73.00
|
17.78
|
|
CKD-EPIcd2
|
70.93
|
0.39
|
21.91
|
76.11
|
17.76
|
70.58
|
0.83
|
22.15
|
73.95
|
18.37
|
|
MDRDmac
|
69.03
|
-1.51
|
22.92
|
73.83
|
18.63
|
69.15
|
-0.60
|
21.67
|
73.54
|
19.25
|
|
Xiangya
|
71.69
|
1.14
|
22.11
|
74.36
|
17.93
|
71.70
|
1.96
|
22.12
|
74.22
|
18.49
|
|
Newly developed equations
|
|
|
|
|
|
|
|
|
|
SC-GFRcr
|
70.22
|
-0.32
|
20.55
|
78.04
|
17.49
|
70.00
|
0.26
|
20.83
|
75.85
|
17.89
|
|
SC-CKD-EPIcr
|
66.82
|
-3.72
|
20.62
|
79.44
|
17.18
|
66.56
|
-3.19
|
20.73
|
78.43
|
17.79
|
|
SC-GFRcr-cys
|
70.12
|
-0.43
|
20.78
|
78.33
|
17.36
|
69.74
|
-0.01
|
20.25
|
76.12
|
17.76
|
|
Abbreviations and definitions: CKD-EPIco, Chinese elderly modified CKD-EPI; CKD-EPIcd2, Chinese 2 type diabetes modified CKD-EPI; MDRDmac, MDRD equation modified by mathematical algorithms; Xiangya, eGFR formular created by Xiangya hospital; SC-CKD-EPIcr, Southeast Chinese modification of Chronic Kidney Disease Epidemiology based on serum creatinine; SC-GFRcr-cys, Southeast Chinese modification of eGFR equation based on serum creatinine and serum cystatin C.
* eGFR, bias, and IQR are given as ml/min per 1.73 m2. To convert GFR from ml/min per 1.73 m2 to ml/s per 1.73 m2, multiply by 0.0167. To convert Scr from m mol/l to mg/dl, divide by 88.4.
† Bias refers to measured GFR minus estimated GFR.
‡ IQR refers to the 25–75th percentile.
§P30 refers to percentage of GFR estimates that are within 30% of measured GFR.
Agreement between new equations and reported equations in CKD evaluation
The r coefficient of mGFR and eGFR calculated by the three new equations were all higher than 0.70 in both development and internal validation datasets (Table 4). When dividing all participants into 5 GFR categories, according to mGFR and eGFR, respectively. Categorical agreement rates between the eGFR and mGFR of the new eGFR equations ranged from 59.40% to 62.33%, the weighted kappa values from 0.55 to 0.58, in development and validation datasets. Among those previous reported eGFR equations, Categorical agreements and weighted kappa were all lower than the new equations, whether in the development dataset or the validation dataset.
Table 4. Agreement between new equations and other equations in CKD evaluation
Equations
|
Development dataset (N=1712)
|
Internal validation dataset (N=737)
|
r coefficienta
|
Categorical agreement
|
Weighted
|
r coefficient
|
Categorical agreement
|
Weighted
|
kappa
|
kappa
|
Previously reported equations
|
|
|
|
|
CKD-EPIco
|
0.77
|
58.12
|
0.57
|
0.77
|
59.70
|
0.54
|
CKD-EPIcd2
|
0.76
|
57.71
|
0.54
|
0.76
|
60.38
|
0.53
|
MDRDmac
|
0.74
|
55.67
|
0.56
|
0.74
|
59.02
|
0.53
|
Xiangya
|
0.76
|
56.19
|
0.50
|
0.76
|
58.75
|
0.49
|
Newly developed equations
|
|
|
|
|
SC-GFRcr
|
0.78
|
59.70
|
0.57
|
0.77
|
61.47
|
0.55
|
SC-CKD-EPIcr
|
0.78
|
59.40
|
0.58
|
0.77
|
62.14
|
0.55
|
SC-GFRcr-cys
|
0.78
|
59.70
|
0.58
|
0.78
|
61.33
|
0.56
|
Abbreviations and definitions: CKD-EPIco, Chinese elderly modified CKD-EPI; CKD-EPIcd2, Chinese 2 type diabetes modified CKD-EPI; MDRDmac, MDRD equation modified by mathematical algorithms; Xiangya, eGFR formular created by Xiangya hospital; SC-CKD-EPIcr, Southeast Chinese modification of Chronic Kidney Disease Epidemiology based on serum creatinine; SC-GFRcr-cys, Southeast Chinese modification of eGFR equation based on serum creatinine and serum cystatin C.
a r coefficient, Spearman’s correlation coefficient.
Table 5. Sensitivity and specificity of eGFR equation in predicting CKD Stages 3–5
Equation
|
Sensitivity (%)
|
Specificity (%)
|
AUC (95% CI)
|
Previous reported equations
|
|
|
|
CKD-EPIco
|
66.24
|
94.27
|
0.80(0.79-0.82)
|
CKD-EPIcd2
|
71.23
|
92.50
|
0.82(0.80-0.83)
|
MDRDmac
|
74.36
|
89.60
|
0.82(0.80-0.83)
|
Xiangya
|
61.48
|
95.78
|
0.77(0.77-0.80)
|
Newly developed equations
|
|
|
|
SC-GFRcr
|
72.16
|
92.63
|
0.82(0.81-0.84)
|
SC-CKD-EPIcr
|
76.33
|
90.04
|
0.83(0.81-0.84)
|
SC-GFRcr-cys
|
72.74
|
92.06
|
0.82(0.81-0.84)
|
Abbreviations and definitions: CKD-EPIco, Chinese elderly modified CKD-EPI; CKD-EPIcd2, Chinese 2 type diabetes modified CKD-EPI; MDRDmac, MDRD equation modified by mathematical algorithms; Xiangya, eGFR formular created by Xiangya hospital; SC-CKD-EPIcr, Southeast Chinese modification of Chronic Kidney Disease Epidemiology based on serum creatinine; SC-GFRcr-cys, Southeast Chinese modification of eGFR equation based on serum creatinine and serum cystatin C.
As shown in Table 5, the AUCs of SC-GFRcr, SC-CKD-EPIcr, SC-GFRcr-cys, CKD-EPIcd2, and MDRDmac ranged from 0.82 to 0.83. While, those of CKD-EPIco and Xiangya equations were 0.80 and 0.77. The sensitivity of SC-CKD-EPIcr was highest (76.33%), followed by MDRDma (74.36%) equation, while others were all lower than 73%. The specificity of SC-GFRcr, SC-CKD-EPIcr and SC-GFRcr-cys equations were 92.63%, 90.04% and 92.06%, respectively, which were lower than those of CKD-EPIco and Xiangya, and higher than those of MDRDmac equation.
Comparing the performance of the new equations in subgroups
We compared the performance of the new equations among different age, sex, diabetes and mGFR categories in the whole population. The accuracy stratified by age was depicted in Supplementary Table S6. The P30% of SC-CKD-EPIcr equations were all highest (> 75%), when comparing with other equations, in all age groups. And it was highest when individuals aged > 60 years (P30 > 80%). CKD estimation was most accurate when using SC-CKD-EPIcr equation for females (P30, 79.68%), and the SC-GFRcr-cys equation estimated renal function most accurately in males (P30, 78.46%) (Supplementary Table S7). Among patients with diabetes (Supplementary Table S8), the SC-CKD-EPIcr equation provided higher accuracy compared to the other formulas (P30, 79.91%). Whereas in individuals without diabetes, higher accuracy was observed for the SC-GFRcr-cys equation (P30, 76.11%). In Supplementary Table S9, when mGFR ≥ 90 ml/min/1.73m2, highest accuracy was observed for the SC-GFRcr (P30, 83.28%), when 60 ≤ mGFR < 90 ml/min/1.73m2,the SC-CKD-EPIcr equation estimated renal function most accurately (P30, 92.37%). Among population with mGFR < 60 ml/min/1.73m2, higher accuracy was observed for the SC-CKD-EPIcr equation (P30, 63.46%) either.