DOI: 10.1111/eci.12344

ORIGINAL ARTICLE Risk score to predict mortality in continuous ambulatory peritoneal dialysis patients Chen Zhao, Qimei Luo, Xi Xia, Feng He, Fenfen Peng, Xueqing Yu and Fengxian Huang Department of Nephrology, The First Affiliated Hospital, Sun Yat-sen University, Key Laboratory of Nephrology, Ministry of Health of China, Guangdong Provincial Key Laboratory of Nephrology, Guangzhou, Guangdong, China

ABSTRACT Background Patients with continuous ambulatory peritoneal dialysis (CAPD) have high all-cause mortality risk that varies extensively among different conditions. The objective of this study was to develop and validate risk models to predict the 2-year all-cause mortality risks of CAPD patients. Material and methods A total of 1354 patients who received CAPD treatment > 3 months from a single dialysis centre were enrolled into the study from January 1, 2006 to December 31, 2011 and followed up until June 30, 2013. The dataset was randomly divided into the derivation dataset (2/3, n = 903) and the validation dataset (1/3, n = 451). Baseline information, including demographic characteristics, comorbid conditions and laboratory data, was recorded and included in the models. Risk models were developed using Cox proportional hazards regression. C-statistic, Akaike Information Criterion, Hosmer-Lemeshow v2 test and net reclassification improvement (NRI) were performed to evaluate model prediction and validation. Results During the entire follow-up period, 175 (19
38%) and 85 (18
85%) patients died in the derivation and validation datasets respectively. A model that included age, diabetes mellitus, hypertension, cardiovascular disease, diastolic blood pressure, serum albumin, serum creatinine, phosphate, haemoglobin and fasting blood glucose demonstrated good discrimination in the derivation and validation datasets to predict 2-year all-cause mortality (C-statistic, 0
790 and 0
759, respectively). In the validation dataset, the above model performed good calibration (v2 = 2
08, P = 0
98) and NRI (7
37% compared with model 2, P = 0
05). Conclusions The risk model can accurately predict 2-year all-cause mortality in Chinese CAPD patients and external validation is needed in future. Keywords All-cause mortality, continuous ambulatory peritoneal dialysis, predictive model, risk score. Eur J Clin Invest 2014; 44 (11): 1095–1103

Introduction Due to many advantages, continuous ambulatory peritoneal dialysis (CAPD) is being increasingly chosen for end-stage renal disease (ESRD) patients, especially in Asian countries [1]. However, the mortality risk of CAPD patients remains high compared with the general population [2,3]. Hence, to delay or prevent death, it is crucial for clinicians to identify early the patients who are at high risk of mortality. Recently, two interesting studies were published. A ‘surprise’ question (‘Would I be surprised if this patient died in the next year?’) was asked to clinicians in haemodialysis or peritoneal dialysis centres, and the results (no vs. yes) were found to help identify patients who were at high risk for short-term mortality [4,5]. The classification was mainly based on the subjective impressions and Chen Zhao and Qimei Luo contributed equally to this study.

experiences of clinicians. However, the question still remains as to how clinicians can make the prediction more objective, sensitive and specific, such as in a mode of scores [6,7]. Several risk factors for mortality in CAPD patients have been identified, for example older age, diabetes mellitus, hypertension, serum albumin, haemoglobin, serum creatinine, etc [8–13], which implies that the mortality risk varies significantly among different patient conditions. A reliable and effective risk scoring model based on these factors is currently needed for predicting mortality risk in CAPD patients. Several dialysis mortality risk prediction methods, some of which were complex and inaccurate in clinical use, have been explored and predominantly focused on haemodialysis patients in white populations [14– 17]. However, there was a difference in mortality risk between haemodialysis and peritoneal dialysis patients [18], and till now, no previous study has established a direct mortality risk model for CAPD patients. The objective of this study was to

European Journal of Clinical Investigation Vol 44

1095

C. ZHAO ET AL.

develop and validate risk models for predicting 2-year all-cause mortality among CAPD patients.

www.ejci-online.com

laboratory variables were assayed before dialysis in the biochemical laboratory of the First Affiliated Hospital of Sun Yat-sen University.

Materials and methods Statistical analysis Participants This was a prospective, observational, cohort study designed to establish risk score models for predicting all-cause mortality in patients with CAPD from a single peritoneal dialysis centre at the First Affiliated Hospital of Sun Yat-sen University, Guangzhou, China. The study population included 1354 patients older than 18 years who had received CAPD for more than 3 months between January 1, 2006 and December 31, 2011. Malignant tumour patients with poor life expectancy and patients with incomplete baseline demographic and clinical information were excluded from this analysis. The only end point of this study was all-cause mortality. Patients were followed up until they died or were censored for kidney transplantation, loss to follow-up, and a transfer to haemodialysis or a nonparticipating dialysis centre. The last date of follow-up was June 30, 2013, and a minimum of 18 months of follow-up data were available. The study patients were followed every 1 or 3 months by dialysis centre clinicians or nurses and the vital status of patients was recorded. The 1354 patients with CAPD were randomly classified into a derivation dataset (2/3, n = 903) and a validation dataset (1/3, n = 451). Ethics approval was granted by the Clinical Research Ethics Committee of the First Affiliated Hospital of Sun Yat-sen University. Written informed consent was obtained from all participants before inclusion.

Predictors Baseline demographic and clinical data were collected from all participants at the initiation of CAPD treatment. These included: age, gender, body mass index (BMI), history of diabetes mellitus (DM), hypertension (HP), cardiovascular disease (CVD), systolic blood pressure (SBP), diastolic blood pressure (DBP), haemoglobin (Hb), serum albumin (Alb), total serum triglycerides (TG), high-density lipoprotein cholesterol (HDLC), low-density lipoprotein cholesterol (LDL-C), high-sensitivity C-reactive protein (hsCRP), calcium (Ca), phosphate (P), serum creatinine (sCr), serum uric acid (UA), fasting blood glucose (FBG) and intact parathyroid hormone (iPTH). The diagnosis of DM at the beginning of CAPD was based on the diagnostic criteria of the American Diabetes Association [19]. The diagnosis of hypertension was based on the history of antihypertensive medication use or two separate blood pressure measurements of SBP > 140 mmHg or DBP > 90 mmHg [20]. CVD was defined as a history of ischaemic heart disease, cerebrovascular disease or peripheral vascular disease [17]. All

1096

Patient demographic and clinical data were presented as frequency (percentage), mean  standard deviation (SD) or median (25th–75th percentile), as appropriate. Baseline differences between the derivation dataset and the validation dataset were tested using the chi-square test, Student’s t-test, or nonparametric Mann–Whitney test, where appropriate. Variables that were associated with all-cause mortality in univariate Cox proportional hazard regression (P < 0
1) were selected as candidate predictors. Combined with clinical guidance on the basis of prior publications, UA and P were also selected as candidate predictors [21,22].The assumption of proportionality of the hazards was met for all models by checking Schoenfeld residuals. A series of predictive mortality models was exclusively established in the derivation dataset, whose variables and b-coefficients were used to calculate the mortality risk in the validation dataset [6,23]. Based on the Cox proportional hazard model, the all-cause mortality risk for CAPD patients over 2 years (t = 2) was estimated as follows: Probability (all-cause mortality) = 1 S(t) ^exp[f(x) – f(xavg)], where f(x) = b1x1 + b2x2 +. . .. . .+ bjxj and f (xavg) = b1x1avg + b2x2avg + . . .. . .+ bjxjavg. In the above model, S (t = 2) is the 2-year survival rate estimated when all risk covariates are at their average values, b1, . . .. . ., bj are the estimated coefficients of baseline covariates from Cox regression, x1, . . .. . ., xj are the values of baseline covariates and x1avg. . .. . . xjavg are the average values of baseline covariates [6,7].Time was set at 2 years for calculating predictive risk, because 63
12% patients in the derivation dataset had a follow-up of more than 2 years. The performance of the models was assessed by discrimination, goodness of fit, reclassification and calibration as follows. Harrel’s C-statistic, which is similar to the area under the receiver operator curve (AUC), was calculated to evaluate the discrimination [24]. Good discrimination, when the C-statistic is above 0
75, indicates the ability of the model to distinguish accurately between two classes of outcomes. Akaike Information Criterion (AIC), which evaluates the best model to interpret the data with the minimal free parameters, was calculated to assess the goodness of fit for any model [25]. A smaller AIC value indicates a better fitting effect. Net reclassification improvement (NRI), which evaluates the correctness of moving patients from one stratum to another as a result of adding new parameters, was calculated to assess the reclassification [26]. A positive NRI value signifies a reclassification improvement of the new risk stratum compared with the old one. The HosmerLemeshow v2 test, which compares the predicted versus observed risk in deciles, was calculated to evaluate the

ª 2014 Stichting European Society for Clinical Investigation Journal Foundation

PREDICT ALL-CAUSE MORTALITY IN CAPD

calibration [27]. A statistically nonsignificant result (P > 0
05) suggests good agreement between predicted and observed probabilities. At different cut-off risks for models, the sensitivity, specificity, Youden’s index, positive predictive values and negative predictive values were calculated. The maximum Youden’s index was used to choose the best cut-off point. Statistical analyses were performed using STATA/SE version 12.0 (Stata Corp, College Station, TX, USA). P < 0
05 was defined as statistical significance for a two-tailed test.

Results Patient clinical characteristics The clinical characteristics of the derivation and validation datasets are shown in Table 1. Patients in the derivation and validation datasets were similar in age, gender, history of DM, HP, CVD, SBP, DBP, Hb, Alb, TG, HDL-C, hsCRP, Ca, P, Scr, UA, FBG, iPTH and dialysis vintage (P > 0
05).Compared with the validation dataset, patients in the derivation dataset had a

Table 1 Clinical baseline characteristics of the derivation and validation datasets Patient characteristics

Derivation dataset (n = 903)

Validation dataset (n = 451)

P-value*

Demographics Age (years)

47
68  15
32

49
02  15
05

0
14

Male, n (%)

522 (57
81)

267 (59
20)

0
64

2

BMI (kg/m )

21
42  3
01

21
78  3
19

0
04

Comorbid conditions, n (%) Diabetes mellitus

236 (26
14)

111 (24
61)

0
55

Hypertension

590 (65
34)

296 (65
63)

0
95

Cardiovascular disease

246 (27
24)

120 (26
61)

0
85

SBP (mmHg)

136
91  20
13

136
43  19
46

0
68

DBP (mmHg)

84
31  14
12

83
41  13
56

0
28

Haemoglobin (g/L)

95
46  18
00

96
26  17
79

0
46

Serum albumin (g/L)

36
32  4
31

36
41  4
67

0
75

2
23  0
19

2
22  0
20

0
68

Blood pressure

Laboratory data

Calcium (mmol/L) Phosphorus (mmol/L)

1
70  0
42

1
68  0
43

0
40

Total triglycerides (mmol/L)

1
73  1
18

1
65  0
96

0
16

HDL-C (mmol/L)

1
17  0
33

1
15  0
32

0
35

LDL-C (mmol/L)

2
88  0
82

2
78  0
79

0
04

hsCRP (mg/L)

3
26 (1
21–6
58)

3
26 (1
32–7
75)

0
12

Uric acid (lmol/L)

421
45  87
44

429
89  94
20

0
11

Serum creatinine (lmol/L)

893
76  298
10

880
57  299
66

0
46

iPTH (pg/mL) FBG (mmol/L) Dialysis vintage (months)

299
03 (174
52–472
81)

290
40 (175
74–466
63)

0
74

5
59  2
06

5
47  1
86

0
32

33
86  19
62

34
31  19
48

0
68

BMI, body mass index; BP, blood pressure; SBP, systolic blood pressure; DBP, diastolic blood pressure; HDL-C, high-density lipoprotein cholesterol; LDL-C, lowdensity lipoprotein cholesterol; hsCRP, high-sensitivity C-reactive protein; iPTH, intact parathyroid hormone; FBG, fasting blood glucose. Continuous variables are presented as means  SD or medians (25th–75th percentile). Categorical variables are presented as frequency (percentage). *P-value is for the difference between patients in the derivation dataset and patients in the validation dataset (t-test or Mann–Whitney test for continuous variables, Chi-square test for categorical variables).

European Journal of Clinical Investigation Vol 44

1097

C. ZHAO ET AL.

www.ejci-online.com

higher level of LDL-C (2
88  0
82 vs. 2
78  0
79 mmol/L, P = 0
04) and a lower BMI (21
42  3
01 vs. 21
78  3
19 kg/m2, P = 0
04). During the entire follow-up period, the proportion of allcause mortality events was similar in the derivation and validation datasets (175 [175/903, 19
38%] vs. 85 [85/451, 18
85%] mortality events, respectively, P = 0
83). In the validation dataset, there were 40 [40/451, 8
87%] mortality events during the 2-year follow-up.

Derivation of the prediction model The univariate Cox analysis for variables of all-cause mortality in the derivation dataset is shown in Table S1. Combined with clinical guidance and univariate analysis, the multivariable predictors included in the models were age, DM, HP, CVD, DBP, Hb, Alb, TG, HDL-C, hsCRP, Ca, P, Scr, UA, FBG and iPTH. The adjusted hazard ratios for the predictive variables and the statistics of the discrimination and goodness of fit for the 2-year mortality risk models are shown in Table 2.

Model 1, which only included demographic characteristics and comorbid conditions (age, DM, HP, CVD, and DBP) resulted in the lowest C-statistic (0
748) and the highest AIC (1838
83). After adding the laboratory variables successively into models 2, 3 and 4, the C-statistic and AIC were gradually improved (0
776, 0
790, and 0
793 for C-statistic, respectively; 1815
75, 1806
02 and 1803
78 for AIC respectively). After adding the variables of Ca, iPTH, hsCRP and HDL-C into model 5, there was no improvement for either C-statistic (0
791 vs. 0
793) or AIC (1808
88 vs. 1803
78) compared with that of model 4. Based on these results, models 2 to 5 were selected as the candidate 2-year mortality risk models to be used for further analysis. The predicted probability of all-cause mortality within 2 years was calculated based on the equations shown in Table S2.

Validation of the prediction model Risk prediction models 2, 3, 4, and 5 with fixed b-coefficients were tested using the validation dataset. The C-statistic was

Table 2 Adjusted hazard ratios and goodness of fit for 2-year mortality in the derivation dataset Models Variable

1

2

3

4

5

Age (per 1 year increase)

1
05

1
04

1
04

1
04

1
05

DM (yes vs. no)

1
86

1
54

1
31

1
32

1
34

HP (yes vs. no)

1
23

1
31

1
33

1
39

1
37

CVD (yes vs. no)

1
67

1
59

1
61

1
57

1
58

DBP (per 1 mmHg increase)

0
99

1
00

1
00

1
00

1
00

Alb (per 1 g/L increase)

0
93

0
96

0
95

0
95

sCr (per 1 lmol/L increase)

0
99

0
99

0
99

0
99

P (per 1 mmol/L increase)

2
20

2
01

1
83

1
68

Hb (per 1 g/L increase)

0
98

0
98

0
98

FBG (per 1 mmol/L increase)

1
10

1
09

1
08

TG (per 1 mmol/L increase)

1
08

1
14

UA (per 1 lmol/L increase)

1
00

1
00

Ca (per 1 mmol/L increase)

0
62

iPTH (per 1 pg/mL increase)

1
00

hsCRP (per 1 mg/L increase)

0
99

HDL-C (per 1 mmol/L increase)

1
42

C-statistic AIC

0
748 1838
83

0
776 1815
75

0
790 1806
02

0
793 1803
78

0
791 1808
88

DM, diabetes mellitus; HP, hypertension; CVD, cardiovascular disease; DBP, diastolic blood pressure; Alb, serum albumin; sCr, serum creatinine; P, phosphate; Hb, haemoglobin; FBG, fasting blood glucose; TG, total serum triglycerides; UA, serum uric acid; Ca, calcium; iPTH, intact parathyroid hormone; hsCRP, highsensitivity C-reactive protein; HDL-C, high-density lipoprotein cholesterol; AIC, Akaike Information Criterion.

1098

ª 2014 Stichting European Society for Clinical Investigation Journal Foundation

PREDICT ALL-CAUSE MORTALITY IN CAPD

0
753 [95% confidence interval (CI) 0
670–0
837] for model 2, 0
759 (95% CI 0
679–0
838) for model 3, 0
761 (95% CI 0
679– 0
842) for model 4, and 0
758 (95% CI 0
674–0
841) for model 5. There was no significant difference in the C-statistic among these models (P > 0
05 for all comparisons). Sensitivity, specificity, Youden’s index, positive predictive value and negative predictive value for different cut-off risk points of these models on the validation dataset are shown in Table S3. At the cut-off point of 7%, models 3, 4 and 5 demonstrated the good sensitivity (77
50%, 72
50% and 82
50%, respectively) and obtained the maximum Youden’s index (0
41, 0
37 and 0
40, respectively). Therefore, when NRI was calculated to evaluate potential reclassification improvement, the 7% risk point was selected as the cut-off value as shown in Table S4. Compared

with model 2, models 3, 4 and 5 had the improvements of 7
37%, 3
10% and 7
01%, respectively, although these were not statistically significant (P ≥ 0
05). Compared with model 3, models 4 and 5 had improvements of 4
27% and 0
35%, respectively, which were also not statistically significant (P > 0
05).According to the deciles of predicted risk in the validation dataset, the calibrations of models 2, 3, 4 and 5 were all good, because the observed and predicted risk of 2-year allcause mortality did not significantly differ by the HosmerLemeshow v2 test (P = 0
71, 0
98, 0
31, and 0
38, respectively), as shown in Fig. 1. Model 3 had the lowest Hosmer-Lemeshow v2 value (2
08) compared with other three models, which meant that the observed and predicted risks were in the closest agreement.

Figure 1 The predicted (black) and observed (grey) mortality probability by deciles of the risk score for models 2, 3, 4 and 5, using the validation dataset. The P-values for the Hosmer-Lemeshow v2 test of the four models were all above 0
05.

European Journal of Clinical Investigation Vol 44

1099

C. ZHAO ET AL.

www.ejci-online.com

Given the above results and clinical convenience, model 3 was chosen as the recommended model. Compared with patients whose risk predictions were below the cut-off point of 7% from model 3, those with risk predictions above 7% had a hazard ratios of 5
34 (95% CI 2
54–11
21, P < 0
001) for 2-year all-cause mortality in the validation dataset as shown in Fig. 2. The risk predictions of 2-year all-cause mortality for hypothetical CAPD patients with different conditions were calculated by applying model 3, as shown in Table 3.

Discussion

Figure 2 Kaplan–Meier survival curves for 2-year all-cause mortality in the validation dataset divided by the cut-off point of the risk from model 3. Upper curve (blue), group with a cut-off point of 7% and below; lower curve (green), group with a cutoff point above 7%.

In this study, we developed and validated four risk models for predicting the risk score of all-cause mortality over a 2-year period among patients who received CAPD treatment for more than 3 months. Models 2, 3, 4 and 5 based on patient demographic characteristics, clinical history and laboratory variables resulted in good discrimination and calibration in the derivation and validation datasets. Using 10 routinely collected variables: age, DM, HP, CVD, DBP, Alb, sCr, P, Hb and FBG, model 3 achieved a C-statistic of 0
790 and 0
759 in the derivation and validation datasets, respectively, and a v2 value of 2
08 (P = 0
978) for the Hosmer-Lemeshow test. Patients in the highest and second highest risk deciles had mean predicted risks of 32
62% and 17
58%, respectively, which were very close to the mean observed risks of 28
57% and 16
67%, respectively. Compared with models 2, 4 and 5, model 3 also showed good

Table 3 Predicted risk of 2-year all-cause mortality for hypothetical patients with different conditions using model 3 Age (years)

DM

HP

CVD

DBP (mmHg)

Alb (g/L)

sCr (lmol/L)

P (mmol/L)

Hb (g/L)

FBG (mmol/L)

Predicted risk (%)

40

No

No

No

80

45

700

1
70

110

5
0

1
60

40

No

Yes

No

80

40

800

1
70

100

6
0

3
07

40

Yes

No

No

80

35

900

1
70

90

7
0

4
40

40

Yes

Yes

Yes

80

30

1000

1
70

80

8
0

13
07

50

No

No

No

80

45

700

1
70

110

5
0

2
46

50

No

Yes

No

80

40

800

1
70

100

6
0

4
70

50

Yes

No

No

80

35

900

1
70

90

7
0

6
71

50

Yes

Yes

Yes

80

30

1000

1
70

80

8
0

19
45

60

No

No

No

80

45

700

1
70

110

5
0

3
77

60

No

Yes

No

80

40

800

1
70

100

6
0

7
16

60

Yes

No

No

80

35

900

1
70

90

7
0

10
17

60

Yes

Yes

Yes

80

30

1000

1
70

80

8
0

28
40

DM, diabetes mellitus; HP, hypertension; CVD, cardiovascular disease; DBP, diastolic blood pressure; Alb, serum albumin; sCr, serum creatinine; P, phosphate; Hb, haemoglobin; FBG, fasting blood glucose.

1100

ª 2014 Stichting European Society for Clinical Investigation Journal Foundation

PREDICT ALL-CAUSE MORTALITY IN CAPD

NRI improvement. Therefore, model 3 is recommended for clinical use. The risk predictors identified by our risk models, such as age, DM, HP, CVD, DBP, Alb, sCr, P, Hb, FBG, TG, UA, Ca, iPTH, hsCRP and HDL-C, were basically consistent with those of previous studies searching for independent risk factors of CAPD [8–13,21,22,28]. However, this present study integrated all of these factors into a single risk model not as separate ones. In recent years, some new independent risk factors have been identified, for example serum troponin T, NT-proBNP, Angiopoietin-2, left atrial volume, etc [29–32]. Our models did not include these new factors because they were not routinely captured for every patient. In the future, these predictors might be incorporated into risk models to obtain more accurate predictions. Several mortality prediction tools have been reported for dialysis patients. Barrett et al. [14] conducted a prospective study with 822 dialysis patients at Canadian centres using age and comorbidity to predict death within 6 months of first dialysis. Unfortunately, the best fit discriminate models still failed to accurately predict death. Miskulin et al. [15] compared four comorbidity instruments in a large US dialysis population using the discriminatory accuracy of 1-year mortality predictions. The Index of Coexistent Diseases (ICED) performed better discrimination (AUC 0
73) than the other three indices and after the addition of race and serum albumin, the discrimination of ICED was improved further (AUC 0
77). But the study was limited by its complexity (166 items) and some predictors which were not routinely collected. Liu et al. [16] developed a new comorbidity index assigned by numerical weights for mortality prediction. However, the index only reached a Cstatistic of 0
669 without considering laboratory measurements. Wagner et al. [17] developed and validated a 3-year mortality model in incident dialysis patients including haemodialysis and peritoneal dialysis patients with a C-statistic of 0
73 in the validation cohort. However, in the cohort whose race proportion was mainly white (83
3%), haemodialysis patients accounted for the major proportion of total participants, and peritoneal dialysis patients accounted for < 30%. Our risk prediction models might play important roles in clinical treatment and research. Table 3 demonstrated a considerable heterogeneity among the predicted risks of 2-year allcause mortality for some hypothetical patients with different conditions. At the cut-off point of 7% for 2-year mortality risk, models 3, 4 and 5 showed good sensitivity and specificity with a maximum Youden’s index. The early evaluation of mortality risk will help clinicians and patients decide the treatment plan, for example intensive treatment or palliative care interventions [4,5]. In models 2, 3, 4 and 5 (as shown in Table S3), negative predictive values were all higher than positive predictive values for 2-year all-cause mortality risk, which might indicate

that lower risk values were more consistent with predicting patients to be alive throughout the 2-year follow-up than higher values predicting those died in the period. In clinical research, our models may be used to determine patients who are at high mortality risk as the study participants, which might reduce the risk-treatment interactions [33,34]. Meanwhile, our study has some limitations. First, the cause of ESRD was not included in the model analysis due to missing data or inaccurate speculation in some patients, despite the fact that it is considered to be a risk predictor [17]. Second, the changes in laboratory variables during treatment, peritoneal dialysis prescription and the use of drugs were all excluded in the model analysis. The reason for the inclusion of only baseline variables was to establish the model for risk stratification and decision-making at the start of CAPD. Third, patients who died in the first three dialysis months, who had high mortality risk, were excluded from the analysis. Finally, external validation patients are required to further validate these models. In summary, we have developed and validated predictive risk models for 2-year all-cause mortality in CAPD patients with good accuracy in discrimination and calibration. Given the increasing use of CAPD, these models may play an important role in identifying the high mortality risk of CAPD patients.

Acknowledgements We thank the trial nurses in our peritoneal dialysis centre. We express our gratitude to all patients who participated in the study. We also thank MPH. Tom Robinson from University of Auckland for sharing the statistical methods.

Funding This study was funded by the grants from the National Key Basic Research Program of China (Grant No. 2011CB504005), the National Key Technology Research and Development Program of the Ministry of Science and Technology of China (Grant No. 2011BAI10B05), the National Natural Science Foundation of China (Grant No. 81170765) and the Guangdong Natural Science Foundation (Grant No. S2011020002359).

Disclosures None. Address Guangdong Provincial Key Laboratory of Nephrology, Department of Nephrology, The First Affiliated Hospital, Sun Yat-sen University, Guangzhou 510080, China (C. Zhao, Q. Luo, X. Xia, F. He, F. Peng, X. Yu, F. Huang). Correspondence to: Fengxian Huang, Department of Nephrology, The First Affiliated Hospital, Sun Yat-sen University, Guangdong Provincial Key Laboratory of Nephrology,

European Journal of Clinical Investigation Vol 44

1101

C. ZHAO ET AL.

58th, Zhongshan Road II, Guangzhou 510080, China. Tel.: +8620 87766335; fax: +86 20 87769673; e-mail: [email protected] Received 15 May 2014; accepted 23 September 2014 References 1 Jain AK, Blake P, Cordy P, Garg AX. Global trends in rates of peritoneal dialysis. J Am Soc Nephrol 2012;23:533–44. 2 Han SH, Lee JE, Kim DK, Moon SJ, Kim HW, Chang JH et al. Longterm clinical outcomes of peritoneal dialysis patients: single center experience from Korea. Perit Dial Int 2008;28(Suppl. 3):S21–6. 3 Nordio M, Limido A, Maggiore U, Nichelatti M, Postorino M, Quintaliani G et al. Survival in patients treated by long-term dialysis compared with the general population. Am J Kidney Dis 2012;59:819– 28. 4 Moss AH, Ganjoo J, Sharma S, Gansor J, Senft S, Weaner B et al. Utility of the ‘surprise’ question to identify dialysis patients with high mortality. Clin J Am Soc Nephrol 2008;3:1379–84. 5 Pang WF, Kwan BC, Chow KM, Leung CB, Li PK, Szeto CC. Predicting 12-month for peritoneal dialysis patients using the ‘surprise’ question. Perit Dial Int 2013;33:60–6. 6 Tangri N, Stevens LA, Griffith J, Tighiouart H, Djurdjev O, Naimark D et al. A predictive model for progression of chronic kidney disease to kidney failure. JAMA 2011;305:1553–9. 7 Elley CR, Robinson T, Moyes SA, Kenealy T, Collins J, Robinson E et al. Derivation and validation of a renal risk score for people with type 2 diabetes. Diabetes Care 2013;36:3113–20. 8 Castrale C, Evans D, Verger C, Fabre E, Aguilera D, Ryckelynck JP et al. Peritoneal dialysis in elderly patients: report from the French Peritoneal Dialysis Registry (RDPLF). Nephrol Dial Transplant 2010;25:255–62. 9 Yang X, Yi C, Liu X, Guo Q, Yang R, Cao P et al. Clinical outcome and risk factors for mortality in Chinese patients with diabetes on peritoneal dialysis: a 5-year clinical cohort study. Diabetes Res Clin Pract 2013;100:354–61. 10 Ortega LM, Materson BJ. Hypertension in peritoneal dialysis patients: epidemiology, pathogenesis, and treatment. J Am Soc Hypertens 2011;5:128–36. 11 Mehrotra R, Duong U, Jiwakanon S, Kovesdy CP, Moran J, Kopple JD et al. Serum albumin as a predictor of mortality in peritoneal dialysis: comparisons with hemodialysis. Am J Kidney Dis 2011;58:418–28. 12 Molnar MZ, Mehrotra R, Duong U, Kovesdy CP, Kalantar-Zadeh K. Association of hemoglobin and survival in peritoneal dialysis patients. Clin J Am Soc Nephrol 2011;6:1973–81. 13 Park J, Mehrotra R, Rhee CM, Molnar MZ, Lukowsky LR, Patel SS et al. Serum creatinine level, a surrogate of muscle mass, predicts mortality in peritoneal dialysis patients. Nephrol Dial Transplant 2013;28:2146–55. 14 Barrett BJ, Parfrey PS, Morgan J, Barre P, Fine A, Goldstein MB et al. Prediction of early death in end-stage renal disease patients starting dialysis. Am J Kidney Dis 1997;29:214–22. 15 Miskulin DC, Martin AA, Brown R, Fink NE, Coresh J, Powe NR et al. Predicting 1 year mortality in an outpatient haemodialysis population: a comparison of comorbidity instruments. Nephrol Dial Transplant 2004;19:413–20. 16 Liu J, Huang Z, Gilbertson DT, Foley RN, Collins AJ. An improved comorbidity index for outcome analyses among dialysis patients. Kidney Int 2010;77:141–51.

1102

www.ejci-online.com

17 Wagner M, Ansell D, Kent DM, Griffith JL, Naimark D, Wanner C et al. Predicting mortality in incident dialysis patients: an analysis of the United Kingdom Renal Registry. Am J Kidney Dis 2011;57:894– 902. 18 Chaudhary K, Sangha H, Khanna R. Peritoneal dialysis first: rationale. Clin J Am Soc Nephrol 2011;6:447–56. 19 American Diabetes Association. Diagnosis and classification of diabetes mellitus. Diabetes Care 2005;28(Suppl. 1):S37–42. 20 He F, Wu X, Xia X, Peng F, Huang F, Yu X. Pneumonia and mortality risk in continuous ambulatory peritoneal dialysis patients with diabetic nephropathy. PLoS ONE 2013;8:e61497. 21 Latif W, Karaboyas A, Tong L, Winchester JF, Arrington CJ, Pisoni RL et al. Uric acid levels and all-cause and cardiovascular mortality in the hemodialysis population. Clin J Am Soc Nephrol 2011;6:2470–7. 22 Palmer SC, Hayen A, Macaskill P, Pellegrini F, Craig JC, Elder GJ et al. Serum levels of phosphorus, parathyroid hormone, and calcium and risks of death and cardiovascular disease in individuals with chronic kidney disease: a systematic review and meta-analysis. JAMA 2011;305:1119–27. 23 Sullivan LM, Massaro JM, D’Agostino RB Sr. Presentation of multivariate data for clinical use: The Framingham Study risk score functions. Stat Med 2004;23:1631–60. 24 Harrell FE Jr, Lee KL, Mark DB. Multivariable prognostic models: issues in developing models, evaluating assumptions and adequacy, and measuring and reducing errors. Stat Med 1996;15:361–87. 25 Deng J, Liang H, Wang D, Sun D, Ding X, Pan Y et al. Enhancement the prediction of postoperative survival in gastric cancer by combining the negative lymph node count with ratio between positive and examined lymph nodes. Ann Surg Oncol 2010;17:1043– 51. 26 Pencina MJ, D’Agostino RB Sr, Steyerberg EW. Extensions of net reclassification improvement calculations to measure usefulness of new biomarkers. Stat Med 2011;30:11–21. 27 Chambless LE, Cummiskey CP, Cui G. Several methods to assess improvement in risk prediction models: extension to survival analysis. Stat Med 2011;30:22–38. 28 Takahashi R, Ito Y, Takahashi H, Ishii H, Kasuga H, Mizuno M et al. Combined values of serum albumin, C-reactive protein and body mass index at dialysis initiation accurately predicts long-term mortality. Am J Nephrol 2012;36:136–43. 29 Havekes B, van Manen JG, Krediet RT, Boeschoten EW, Vandenbroucke JP, Dekker FW et al. Serum troponin T concentration as a predictor of mortality in hemodialysis and peritoneal dialysis patients. Am J Kidney Dis 2006;47:823–9. 30 Paniagua R, Ventura MD, Avila-Dıaz M, Hinojosa-Heredia H, Mendez-Duran A, Cueto-Manzano A et al. NT-proBNP, fluid volume overload and dialysis modality are independent predictors of mortality in ESRD patients. Nephrol Dial Transplant 2010;25:551–7. 31 David S, John SG, Jefferies HJ, Sigrist MK, K€ umpers P, Kielstein JT et al. Angiopoietin-2 levels predict mortality in CKD patients. Nephrol Dial Transplant 2012;27:1867–72. 32 Kim SJ, Han SH, Park JT, Kim JK, Oh HJ, Yoo DE et al. Left atrial volume is an independent predictor of mortality in CAPD patients. Nephrol Dial Transplant 2011;26:3732–9. 33 Wang R, Lagakos SW, Ware JH, Hunter DJ, Drazen JM. Statistics in medicine–reporting of subgroup analyses in clinical trials. N Engl J Med 2007;357:2189–94. 34 Kent DM, Hayward RA. Limitations of applying summary results of clinical trials to individual patients: the need for risk stratification. JAMA 2007;298:1209–12.

ª 2014 Stichting European Society for Clinical Investigation Journal Foundation

PREDICT ALL-CAUSE MORTALITY IN CAPD

Supporting Information Additional Supporting Information may be found in the online version of this article: Table S1. Univariate Cox analysis of risk factors for all-cause mortality.

Table S2. Coefficients for the derivation dataset 2-year all-cause mortality risk equations of models 2, 3, 4, and 5. Table S3. The predictive performance for different cutoff risk points of models 2, 3, 4, and 5 in the validation dataset. Table S4. Net Reclassification Improvement of models 2, 3, 4, and 5 in the validation dataset.

European Journal of Clinical Investigation Vol 44

1103

Copyright of European Journal of Clinical Investigation is the property of Wiley-Blackwell and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use.