Prognostic factors of all-cause mortalities in continuous ambulatory peritoneal dialysis: a cohort study

Background The role of small solute clearance on mortalities in patients with CAPD has been controversial. We therefore conducted a study with 3 years' follow up in adult patients who participated in the CAPD-first policy. Methods There were 11,523 patients with end-stage renal disease who participated in the CAPD-first policy between 2008 and 2011. Among them, 1,177 patients were included in the retrospective cohort study. A receiver operating characteristic curve was applied to calibrate the cutoffs of tKt/V, rKt/V and tCrcl. Kaplan-Meier and Cox-regression models with time varying covariates were applied to estimate overall death rate, probability of death and prognosis, respectively. Results The cutoffs of rKt/V and tKt/V were 0.25 and 1.75, respectively. The Cox regression suggested that the higher these clearance parameters, the lower the risks of death after adjusting for covariables. The risks of death for those above these cutoffs were 57% (HR = 0.43, 95% CI: 0.31, 0.60) and 29% (HR = 0.71, 95% CI: 0.52, 0.98) lower for rKt/V and tKt/V, respectively. Age, serum albumin, hemoglobin, systolic blood pressure, and ultra-filtration volume significantly affected the mortality outcome. Conclusions Our study suggested that the cutoffs of 0.25 and 1.75 for rKt/V and tKt/V might be associated with mortality in CAPD patients. A minimum tKt/V of 1.75 should be targeted, but increased dialysis dosage to achieve tKt/V > 2.19 adds no further benefit. Serum albumin, hemoglobin, SBP, and UF volume are also associated with mortality. However, our study may face with selection and other unobserved confounders, so further randomized controlled trials are required to confirm these cutoffs.


Background
The global burden of end-stage kidney disease (ESRD) is increasing rapidly [1][2][3] with a prevalence of 0.2% in the USA, and 0.3% in Thailand respectively [4,5]. Renal replacement therapy (i.e., renal transplantation or dialysis) is required for ESRD patients, and 2 modalities, hemodialysis and continuous ambulatory peritoneal dialysis (CAPD), have been widely used. The numbers of patients on CAPD has been growing rapidly in Asian countries, representing about 80% of dialysis patients in Hong Kong, and 50% in Thailand in 2011. A health care reform scheme for all Thai citizens, called the "Universal Coverage" scheme (UC) was first initiated in 2002, and it has covered renal replacement therapy for CAPD-first treatment since 2008. The usage of four 2-L daily exchanges with double-bag disconnected systems has been a standard CAPD regime in Thailand.
We therefore conducted a cohort study with 3 years' follow up to answer the following research questions: Can tKt/V, rKt/V, and tCrcl significantly predict disease prognosis in CAPD patients? If so, what are the cutoffs that can be used to predict patients' outcomes?

Setting & participants
This cohort study covered patients receiving four 2-L CAPD exchanges at 82 general hospitals, belonging to the Ministry of Public Health, Thailand. Data from the National Health Security Office (NHSO) were retrieved between January 2008 and April 2011. The study was approved by the Ramathibodi Hospital Ethical Committees.
Patients aged 15 years or older were eligible if they met the following criteria: Firstly initiated CAPD and participated in the CAPD first-policy from January 2008 to April 2011, survived more than 1 month after initiating CAPD, and had at least 1tKt/V during the studied period.

Clinical endpoint
The primary outcome of the study was time since first initiation of CAPD therapy to death. Patients were censored if they were lost to follow up, or survived at the end of the study (May 2011). Death referred to all cause mortality and the data were validated by cross-referencing with the death certificate database from the Ministry of the Interior.

Prognostic factors
The studied prognostic factors were renal (i.e., rKt/V and rCrcl) and peritoneal (i.e., pKt/V and pCrcl) small solute clearances. The former clearance was estimated using a ratio of concurrent urea/creatinine excretion in 24-hour urine whereas the latter clearance was estimated using 24-hour dialysate effluent. The total small solute clearance (e.g., tKt/V or tCrcl) was the summation of renal and peritoneal small solute clearance. Urea distribution volume (V) was measured using Watson's formula [16]. These small solute clearances (i.e., rKt/V, tKt/V, tCrcl) and other variables were considered in the prognostic model, which included age, gender, body mass index (BMI), serum albumin, hemoglobin, UF volume, SBP, DBP, and co-morbidities (i.e., diabetes, hypertension and/or cardio-vascular disease (CVD)). Among these variables, tKt/V, rKt/V, tCrcl, BMI, serum albumin, hemoglobin, UF volume, SBP, and DBP were considered in the analysis as time-varying covariates, whereas the rest were fixed variables.

Statistical analyses Imputation
Among 1,177 eligible patients, the data for UF volume, hemoglobin, serum albumin, rKt/V, BMI, tCrcl, and comorbidities were missing in 0.3%, 1.1%, 3.3%, 6.5%, 11.5%, 16.7%, and 36.3% of patients, respectively. For each patient, the last observed value was carried forward to replace missing data. Then, the rest of the missing data were imputed using multivariate chain equations. A simulation-based procedure [17,18] with the assumption that data were missing at random was applied. Logistic and linear regressions were applied to predict missing data for dichotomous and continuous data, respectively. Twenty imputations were performed to allow for the uncertainty of imputed data and the summarized values were then used [19].

Calibration of the cutoff threshold
The received operative characteristic (ROC) curve was used to determine the cutoff of tKt/V, rKt/V, and tCrcl that could discriminate death from living patients. Each variable was fitted in the ROC model as both continuous and categorized variables, separately. For categorical variables, they were categorized according to tertile distributions. The likelihood ratio positive (LR+) and Youden's index (i.e., highest sensitivity + specificity −1) [20] were then used to select the cutoff threshold. The performance of the cutoff thresholds suggested by the two methods were incorporated which led to the final cutoffs.
Kaplan-Meier was applied to estimate the overall death rate and probability of death at 12-, 24-, and 36-months after first-initiated CAPD. A patient was censored if s/he had one of the following events: loss to follow-up, withdrew from the CAPD program, or alive at the end of follow up period. Cox-regression models with time varying covariates were applied to assess prognostic effects by fitting equations that separately contained rKt/V and tKt/V. Prognostic scores were then calculated from the final Cox models. The ROC curve analysis was used to estimate the area under the ROC curves (AUC) between the two prognostic scores [21].
All analyses were done using STATA software version 12. The p-value of less than 0.05 was considered statistically significant.
Kaplan-Meier curves were plotted by tKt/V groups (see Figure 2a). This suggested that the failure curves of tKt/V of 1.75-2.19 and > 2.19 groups were similar, but they were lower when compared with tKt/V <  19 when compared to tKt/V < 1.75, respectively. These two tKt/V groups were then combined because their HRs were similar.
The estimated probability of death was decreased as the rKt/V increased (see Figure 2b).  Table 2).
Higher BMI were associated with lower mortality with the adjusted HR of 0.73 (95% CI: 0.53, 1.04) for BMI 21 kg/m 2 or higher when compared with BMI less than 21 kg/m 2 , but this was not statistically significant. The presence of diabetes increased the risk of death by 43% with the adjusted HR of 1.43 (95% CI: 0.98, 2.09), but this was also not statistically significant.

Discussion
We conducted a retrospective cohort study of firstinitiated CAPD patients with a median follow-up time of 22.9 months. The 12-, 24-and 36-months probabilities of death were 6.7% (95% CI: 5.3%, 8.4%), 19.4% (95% CI: 16.7%, 22.4%), and 29.2% (95% CI: 24.7%, 34.4%), respectively with the overall death rate of 9.9 (95% CI: 8.6, 11.4) per 100 patient-years. Prognostic factors of death were explored suggesting that rKt/V was the strongest predictor of death followed by tKt/V. The cutoff thresholds of these parameters that discriminated poor from good prognosis were ≥ 0.25 for rKt/V and ≥ 1.75 for tKt/V. Patients with the rKt/V and tKt/V higher than the cutoffs had respectively 65% and 29% lower risk of death than patients with lower levels than the cutoffs. The mortality rate for our cohort was quite low (i.e., 6.7% and 19.4% for 1-and 2-year probabilities of death) when compared with the mortality rates in other Asian countries, e.g., Indian (20% and 40% ) [22], and China (10% and 21%) [23]. This might be explained by our study included patients who were firstly initiated CAPD while those studies included both prevalent and incident CAPD cases. Modalities of CAPD exchanges were different, we used four 2-L CAPD exchanges whereas those studied used three 2-L CAPD exchanges. Although our study had higher prevalence of diabetes compared with the Chinese study [19], our patients were younger (mean age 50 vs. 54 years) and about three times lower CVD (4.4% vs. 14.2%) than the Chinese study. We included only patients who had tKt/V data and survived at least 1 month after first-initiated CAPD. As such, our mortality rate was lower than previous studies.
A target small solute clearance has long been a critical issue. Since increasing the peritoneal dialysis prescription may increase the patients' discomfort, decrease the quality of life, and harm by metabolic glucose load, the advantages from increased dialysis dose and potential harms should be well-balanced. Unfortunately, the optimum target clearance has not been well-formulated due to lack of previous evidence and our study should be able to fill information for this issue. Both continuous and categorical variables for small solute clearances were explored in the ROC curve analyses and suggested the optimum cutoffs for tKt/V and rKt/V. These cutoffs  were similar to the current tKt/V target suggested by the Kidney Disease Outcome Quality Initiative (KDOQI) guideline [25]. Two randomized controlled trials [13,14] have assessed the effects of tKt/V targets advocated by the previous KDOQI guideline. The ADEMEX study [13] compared the modified prescriptions to achieve a target tCrcl of > 60 L/week/1.73 m 2 with the four 2-L daily exchanges of standard PD solution. They found no significant differences in patient survival, technique survival, and quality of life between intervention and control groups. The Hong Kong study [14] which had randomized patients to 3 different tKt/V targets also showed negative results. Some studies [7,11,12] considered two or more indices of small solute clearance simultaneously in the regression models. Since tKt/V was the sum of pKt/V plus rKt/V, the two variables were highly correlated with the estimated correlation coefficient of 0.388 (p-value < 0.001). Although including them simultaneously in the same model did not cause much multicolinearity as we were aware, adding tKt/V in the model which already contained rKt/V did not improve the explanation of Increased serum albumin, hemoglobin, UF volume, and SBP > 140 mmHg were also associated with a lower risk of mortality. Interestingly, we found an inverse relationship between high SBP and risk of death. This finding is consistent with recent findings in chronic kidney disease/dialysis patients which suggested that low blood pressure could be related to coexisting co-morbidities such as heart failure, autonomic neuropathy, decreased food intake, and worse nutritional status [26][27][28][29]. These explanations may partly be attributable to these paradoxical findings, but the mechanism underlying this reverse relationship is not clearly understood.
Our study has some strengths. Our study is a large CAPD cohort with a median follow up time of nearly 2 years. We properly calibrated the rKt/V and tKt/V cutoffs using ROC curve analysis. We considered study factors (i.e., rKt/V, tKt/V, and tCrcl ) and other observed prognostic factors as time-varying co-variables in the survival analyses, which should be better in explaining disease prognosis than using only single measurements. However, our study also has some limitations. First, the study was not a randomized controlled trial and thus selection and unobserved confounding biases could not be avoided. Second, patients were only eligible if they had data for tKt/V, which was about 10% of the entire cohort; so the representativeness of the whole CAPD patients was limited. Third, we might face with survival bias because of our inclusion criteria and differences of characteristics of eligible versus ineligible patients. Our results may be generalizable to patients who firstly received CAPD with stable conditions, not patients with critical conditions. The specific etiologies of death including their prognostic factors were not evaluated due to lack of information. Among eligible patients, some variables were missing and thus multiple imputation using regression analysis was applied to predict missing values. Twenty imputations were applied taking into account the between-imputation variability.

Conclusion
The cutoffs of 0.25 and 1.75 for rKt/V and tKt/V might be targeted to prevent mortality in CAPD patients. Serum albumin, hemoglobin, SBP, and UF volume are also associated with mortality. However, our study may face with selection and other unobserved confounders. Further randomized control trial is required to confirm these cutoffs.