This study was based on the

N

S

A

R

F

The demographic data, comorbid diseases, types of surgery and RRT, as well as the indications for RRT were documented. The biochemistry data such as complete blood cell count, blood urea nitrogen (BUN), serum creatinine (sCr), glomerular filtration rate (GFR), serum albumin, and serum potassium (sK⁺) were recorded upon ICU admission and RRT initiation. Severity scores including Glascow Coma Scale (GCS) score, Acute Physiology and Chronic Health Evaluation II (APACHE II) 19 score, and Sequential Organ Failure Assessment (SOFA) 20 score were also measured at the two time points. Also, the need for mechanical ventilation was recorded and the usage of inotropic equivalent was calculated to evaluate the vasopressor dose 21. Then we measured and recorded patients' outcome including in-hospital mortality and RRT wean-off.

Definitions were made as following: diabetes, previous usage of insulin or oral hypoglycemic agents; hypertension, blood pressure above 140/90 mmHg or usage of anti-hypertension agents; cardiac failure, low cardiac output with a central venous pressure (CVP) above 12 mmHg and an dopamine equivalent above 5 μg/kg/min 21; chronic kidney disease (CKD), sCr of 1.5 mg/dl or greater documented prior to this admission 22; sepsis, persisted or progressive signs and symptoms of the systemic inflammatory response syndrome with a documented or presumed persistence of infection 23; RRT wean-off, cessation from RRT for at least 30 days 15.

The types of major abdominal surgery were further divided into five categories depending on the involvement of abdominal organs: (1) hepatobiliary organ, (2) upper gastrointestinal (GI) tract, (3) lower GI tract, (4) urological organs, and (5) other sites. 'Upper GI tract' was defined as the duodenum and above, while 'lower GI tract' included the area from the jejunum to rectum. If the surgery didn't involve the one of the four major organs (1 to 4), it would be categorized as 'other sites' (5).

The modality of RRT was chosen according to the hemodynamics of the patients. Continuous venovenous hemofiltration was performed, if more than 15 points of inotropic equivalent 15 was required to maintain systemic blood pressure up to 120 mmHg, using high-flux filters (Hemofilter, PAN-10, Asahi Kasei, Japan) and HF 400 (Informed, Geneva, Switzerland). The hemofiltration flow and blood flow blood flow were 35 ml/kg/hour and 200 ml/min, respectively. Replacement fluid was bicarbonate-buffered and was administered predilutionally at a dynamically adjusted rate to achieve the desired fluid therapy goals. Default composition was sodium 142 mEq/l, bicarbonate 33 mEq/l, calcium 2.6 mEq/l, and magnesium 1.4 mEq/l. Intermittent hemodialysis was performed for four hours except for the first and second sessions with a dialysate flow of 500 ml/min and blood flow of 200 ml/min 18, using low-flux polysulfone hemofilters (KF-18C, Kawasumi Laboratories, Shinagawa-ku, Tokyo, Japan). Double lumen catheters were placed as vascular access.

In the ICUs, the indications for RRT initiation were: (1) azotemia (BUN > 80 mg/dL and sCr > 2 mg/dl) with uremic symptoms (encephalopathy, nausea, vomiting, etc); (2) oliguria (urine amount < 200 ml/8 hours) or anuria refractory to diuretics; (3) fluid overload refractory to diuretics use with a CVP level above 12 mmHg or pulmonary edema with a partial pressure of arterial oxygen/fraction of inspired oxygen ratio below 300 mmHg; (4) hyperkalemia (sK⁺ > 5.5 mmol/L) refractory to medical treatment; and (5) metabolic acidosis (a pH < 7.2 in arterial blood gas) 1318. We recorded all the indications of patients upon RRT initiation.

Patients were categorized into two groups (early dialysis (ED) and late dialysis (LD)) according to their RIFLE (Risk, Injury, Failure, Loss, and End stage) classification 24 (Table 1) before RRT initiation. The RIFLE classification was first proposed by the Acute Dialysis Quality Initiative group in an attempt to standardize AKI study, and the scores could be used to predict the mortality after major surgery 2526. There were many studies comparing the prognoses among patients in different categories of RIFLE classification, but only a few studies 2728 compared the outcome among patients who initiated RRT in different categories of RIFLE classification. As in previous studies 272930, we used 'simplified' RIFLE (sRIFLE) classification with only GFR criterion applied for classification because the eight-hourly urine volumes in our database could not match the 6- or 12-hourly urine output criterion in RIFLE classification. Those who initiated RRT when in sRIFLE-R (risk) or sRIFLE-0 26, which means not yet reaching the sRIFLE-R level were defined as 'ED', while in sRIFLE-I (injury) or sRIFLE-F (failure) were classified as 'LD'. The baseline sCr was the data obtained at hospital discharge from the previous admission in those who had more than one admission 29, or the data estimated using the Modification of Diet in Renal Disease (MDRD) equation 31 in those with only one admission (assuming an average GFR of 75 ml/min/1.73 m²). The peak sCr was defined as the highest sCr before RRT initiation in ICUs. The GFR were estimated using the isotope dilution mass spectrometry--traceable four-variable MDRD equation 31.

The endpoint of this study was in-hospital mortality. The survival period was calculated from RRT initiation to mortality (in non-survivors) or hospital discharge (in survivors).

Statistical analyses were performed with the Scientific Package for Social Science for Windows (SPSS, version 13.0, SPSS Inc, Chicago, IL, USA). Continuous data were expressed as mean ± standard deviation unless otherwise specified. Percentage was calculated for categorical variables. Student's t test was used to compare the means of continuous data, whereas Chi-squared test or Fisher's exact test was used to analyze categorical proportions. Then we used backward stepwise likelihood ratio model of Cox proportional hazard method to analyze the independent predictors for in-hospital mortality. The independent variables were selected for multivariate analysis if they had a P ≤ 0.1 on univariate analysis. The basic model-fitting techniques for (1) variable selection, (2) goodness-of-fit assessment, and (3) regression diagnostics (e.g., residual analysis, detection of influential cases, and check for multicollinearity) were used in our regression analyses to ensure the quality of analysis results. Specifically, we used the stepwise variable selection procedure with both significance level for entry and significance level for stay set to 0.15 or larger to select the relevant covariates into the final Cox proportional hazards model. Also, we did an additional analysis adjusting for three clinical relevant variables (namely, sepsis before RRT, mechanical ventilation, and diabetes) regardless of P value because they were considered important. Furthermore, we did the analysis comparing sRIFLE categories against each other for the relative risk (RR) for in-hospital mortality. In statistical testing, two-sided P value less than 0.05 was considered statistically significant.

Finally, Kaplan-Meier survival curves with log-rank test was drawn to express the differences of patient survival between the two groups (ED versus LD).

The RIFLE classification 24 was proposed to standardize the severity of AKI, and it's predictive value for patient outcome was supported by many studies 252632. The stratification about the timing of RRT initiation by RIFLE classification has been recommended by the Acute Kidney Injury Network 33. Our work is among the first few studies examining the relation between prognosis and timing of RRT initiation. We found that late initiation of RRT as defined by 'sRIFLE-I' and 'sRIFLE-F' is an independent predictor for in-hospital mortality in a relative homogenous group of patients with AKI after major abdominal surgery.

Current practice suggests that RRT is indicated for a patient with an abruptly decreased renal function along with clinically significant solute imbalance or volume overload, yet there is no consensus on the definite indication for RRT in terms of any single metabolic or clinical parameters or RIFLE staging 33. Although the benefit of early RRT initiation on survival outcome was revealed by a recent systemic review and meta-analysis 9, the question of 'how early is early enough?' is still unanswered because the early versus late RRT were defined by variable cutoff values of various metabolic parameters such as nitrogenous waste products, sCr, sK⁺ 34, urine amount, or even clinical judgment alone 935. The present study defines the timing of RRT initiation by using RIFLE classification because this has been extensively validated to standardize the severity of AKI 33.

As it is reasonable that the patient survival is artificially extended if it is measured at an earlier time point with better residual renal function and less severity scores, and the so-called survival benefit from early RRT could be accounted for by lead-time bias 34. However, the period from hospital admission to RRT initiation, as well as the severity scores including APACHE II score and SOFA score and almost all clinical parameters upon RRT initiation, which was taken as a starting point to calculate survival period, were of no statistical differences between ED and LD groups (Table 2). Therefore, the argument of lead-time bias would be minimized in the current study.

Two recent published studies 2728 have evaluated the association between the timing of RRT initiation by the RIFLE classification and outcome. Neither of them propose clearly defined indications for RRT. Only 33% patients in one study 28 and none in the other 27 were categorized using both GFR and urine output criteria. The retrospective observational study by Li and colleagues 27 enrolled 106 critical AKI patients treated with continuous RRT. It found that the RIFLE classification may be used to predict 90-day survival after RRT initiation, and further analysis revealed that patient in RIFLE-F had a RR of 1.96 (95% CI: 1.06-3.62) comparing with those in RIFLE-R. The predictive effect was also seen in our work in which the RR of sRIFLE-F to sRIFLE-R was 3.194 (P = 0.014). As to the study of Maccariello and colleagues 28, a prospective cohort study including 214 AKI patients who underwent RRT, the RIFLE classification didn't show discrimination of prognosis in all patient populations. However, the association between RIFLE-F and increased in-hospital mortality was found while conducting a separate analysis study using only patients who underwent ventilation and vasopressors.

In our study, RRT was provided to the patients according to the five criteria, namely, (1) azotemia with uremic symptoms, (2) oliguria or anuria, (3) fluid overload, (4) hyperkalemia, and (5) metabolic acidosis. Although the criteria for RRT were not too loose compared with those in other studies 933, about half of the patients who underwent RRT (n = 51, 52.0%) were categorized into the ED group. This result was not surprising when compared with the largest study on the epidemiology of AKI during the entire ICU stay by Ostermann and Chang 30. Among the total 1847 patients who underwent RRT in that study, only 573 (31.0%) fulfilled the sCr criterion and 691 (37.4%) would probably fulfill the urine criterion for AKI stage III, and the remaining 583 (31.6%) would be classified into earlier stage 36. Actually, RIFLE classification and our own criteria for RRT are different scoring systems. The numbers of our indications for RRT are more than the parameters used in the RIFLE classification (only sCr level, GFR, and urine amount). Although the parameters in the RIFLE classification seems similar to the former two of our five RRT indications, the percentage change in sCr or GFR in RIFLE classification was different from the absolute BUN or sCr level in ours. Furthermore, 'oliguria or anuria' played a significant role as an indication for RRT in our study (45.1% and 36.2% in ED and LD, respectively) (Table 2), but the urine criterion of RIFLE classification was not used in categorizing patients. It means that those who met our study indications and received RRT accordingly may not be considered serious by RIFLE classification.

In critically ill patients, AKI is usually associated with multiple-organ failure. Preventing further renal damage and recovering renal function are largely dependent on recovery of other organ function. Thus, the concept has changed from 'renal replacement' to 'renal support' in ICU patients 373839. However, RRT has often been applied too late 40, leading to prolonged and poorly controlled uremia, restricted nutrition, acidosis, and volume overload 41. In this study, the indications for RRT were not statistically different between ED and LD groups (Table 2), and survivors and non-survivors (detail not shown in the text). Thus, the survival benefit could not be simply explained by the causes of RRT initiation (such as fluid management or toxin removal), and the importance of early initiation of RRT clearly speaks for itself in this study 9.

More than half of patients who underwent RRT following major abdominal surgery died during hospital admission, which is comparable with previous studies 294243. Our study found that LD defined by sRIFLE classification, along with old age, cardiac failure, and pre-RRT SOFA scores, are strong predictors for in-hospital mortality. Old age has been a well-recognized predictor for mortality in critically ill surgical patients in many studies 284344. Cardiac failure is not only characterized by a high rate of hospital readmission and mortality in the general population 45, but is also considered an independent predictors for mortality in critically ill surgical patient with AKI 43. Besides, SOFA score was chosen as a representation of severity score for Cox analysis in our study. The predictive value for poor prognosis in AKI of SOFA score has been reported in other studies 130 as well.

Similar to the report of a systemic review and meta-analysis summarizing all studies published before 2008 9, our data supported the survival benefit in earlier initiation of RRT. However, discordant results existed. Bagshaw and colleagues 46 designed a prospective multicenter observational study enrolling 1238 patients to evaluate the relation between timing of RRT initiation in severe AKI and prognoses. Timing of RRT was assessed by several approaches such as median value and median change of BUN and sCr, and the period from ICU admission to start of RRT. Contrary to our findings, they found late RRT stratified by median sCr was associated with lower mortality. Previous studies 47 using sCr criterion to define early RRT also failed to show survival benefit. The main plausible explanation is that low sCr levels might not necessarily represent a better residual renal function. In contrast, the low sCr could be a marker of reduced muscle mass and malnutrition, and it may be a surrogate marker of volume overload, which in turn might contribute to poor survival 3346.

However, this bias did not exist in our study because the sCr and albumin level were not statistically different between ED and LD groups upon ICU admission and before RRT initiation (Table 2). In fact, the relation between sCr and mortality was ever documented to be paradoxical in dialysis patients, which is called 'reverse epidemiology'. It refers to paradoxical and counter-intuitive epidemiologic associations between survival outcomes and traditional risk factors such as creatinine 48.

It is worthy of mention that the LD group in our study has better baseline renal function (less CKD proportion, lower baseline sCr, higher baseline GFR) but worse pre-RRT renal function. There is no doubt that a larger sCr increase or GFR decrease categorized patients into LD group, but it also gave a hint that those with more sever renal function deterioration have poorer outcome. Actually, both the proportional change of sCr or GFR in RIFLE classification, and the absolute sCr level in the SOFA scores could predict prognoses in our patients. This finding was supported by Coca and colleagues 49 who had disclosed the prognostic importance of a small acute change in sCr in absolute level as well as percentage changes.

Several limitations for this study should be recognized. First, the limited patient number may not be large enough to determine other risk factors for in-hospital mortality. Second, only GFR criterion of RIFLE classification was used in the current study. Although several studies 2930 did the same, it is a shortcoming to lack urine output when applying the RIFLE category. Thus we used the term 'sRIFLE' in our manuscript to distinguish from the original RIFLE. Therefore, observations accrued here might not be extrapolated to patients with AKI elsewhere. Further multicenter randomized clinical trials are warranted to confirm our findings.