We conducted an observational study in a multiple-center database (OUTCOMEREA) from January 1997 to June 2009. Methods of data collection and quality of the database have been described in details elsewhere 10. Briefly, a large set of data on a random sample of patients older than 16 years with ICU stays longer than 24 h was prospectively collected by the senior physicians of the participating ICUs and entered into the database each year. The quality control procedure involved multiple automatic checking of internal consistency and biennial audits.

In accordance with French law, the OUTCOMEREA database was declared to the Commission Nationale de l'Informatique et des Libertés. The study was approved by the ethics committee of Clermont-Ferrand, France. Since the study did not modify patients' management and data were processed anonymously, the need for informed consent was waived.

All patients in the database were eligible. Exclusion criteria were: chronic kidney disease (CKD) (with or without complete loss of kidney function), pre-renal cause of renal dysfunction (that is rapidly reversible functional renal failure), multiple ICU stays, decision to withhold or withdraw life-sustaining treatments, and renal replacement therapy for extra-renal indications (such as, intoxications or cardiogenic shock). CKD was defined either according to the Acute Physiology and Chronic Health Evaluation (APACHE) II definition or a specific code in the database when not requiring dialysis. Pre-renal cause of renal dysfunction was also identified through a specific code in the database. The reason for excluding these patients was that their prognosis may be different from that of patients with prior normal renal function who present a non-rapidly reversible cause of AKI. Patients with multiple ICU stays, or with a decision to withhold or withdraw life-sustaining treatments, were also excluded to avoid confusion in the assessment of hospital mortality.

Among the remaining patients, those in whom AKI occurred were analyzed. AKI was defined according to the RIFLE (Risk, Injury, Failure, Loss and End-stage renal failure) criteria 11, and patients were classified according to the maximum RIFLE class (Risk, Injury or Failure) reached during their ICU stay. The maximum RIFLE class was determined before RRT initiation in patients who received RRT and whenever during the ICU stay in patients who did not. Since the 6- and 12-h urine outputs were not recorded in the database, we used the glomerular filtration rate (GFR) only. The GFR criteria were determined according to changes in serum creatinine from baseline values. As AKI may be present on ICU admission in a high proportion of patients, we chose to assess baseline creatinine values using the Modification of Diet in Renal Disease (MDRD) equation. As recommended by the Acute Dialysis Quality Initiative Group, a normal GFR of 75 ml/min/1.73 m² before ICU admission was assumed 11.

RRT consisted of intermittent hemodialysis or continuous veno-venous hemofiltration/hemodiafiltration. All participating centers were able to provide both techniques of RRT. The decision to start RRT was left at the discretion of the attending ICU physicians.

The following data were recorded:

• baseline characteristics on ICU admission: age, sex, McCabe class (class 1, no fatal underlying disease; class 2, underlying disease fatal within five years; class 3, underlying disease fatal within one year), Simplified Acute Physiology Score (SAPS) II, comorbidities assessed according to the APACHE II definitions, transfer from ward (defined as a stay in an acute-bed ward ≥24 hrs immediately before ICU admission), and admission category (medical, scheduled surgery, or unscheduled surgery),

• during the ICU stay: daily biological parameters (blood urea nitrogen, serum creatinine, kaliemia), daily urine output, daily weight, time from admission to maximum RIFLE class, time to RRT, daily Sequential Organ Failure Assessment (SOFA) score, and modified SOFA score (mSOFA, SOFA - specific renal component),

• on ICU discharge: renal status (recovery or need for prolonged renal support), and length of ICU stay, and,

• on hospital discharge: vital status.

The primary endpoint was hospital mortality.

The secondary endpoints were the length of ICU stay, and renal status on ICU discharge.

Since RRT was not randomly assigned in our study population, treatment selection bias was accounted for by using the propensity technique. When building a propensity score, the main risk is the omission of an important variable in the propensity regression. Thus, we fitted and compared two different models to strengthen our analysis. As recommended, propensity scores were determined through multivariate logistic regression 12, in which RRT was the dependent variable. Independent variables were related to the probability of receiving the treatment and also outcome in order to reduce both the bias and the variance in the estimation of treatment effect 1314. Independent variables introduced in model 1 were: rising creatinine reflected by maximum RIFLE class, oliguria reflected by the 24-h urine output on reaching maximum RIFLE class, and SAPS II score. Independent variables introduced in model 2 were: blood urea nitrogen, serum creatinine and kaliemia measured on reaching maximum RIFLE class (that is, before RRT was started), fluid accumulation (reflected by the difference between patients' weight recorded on reaching maximum RIFLE class and that recorded on ICU admission), and SAPS II score.

Using an algorithm 15, we matched patients who received RRT during their ICU stay to other AKI patients who did not on the basis of each of the two propensity scores that we built (model 1 and model 2). Specifically, we sought to match each patient with RRT up to three controls who had the closest propensity score (within 0.05 on a scale of 0 to 1).

Besides, patients were also matched on center and period of admission to account for possible inconsistent institutional practices or changes in RRT practices over time. Age (+/- 5 years) was the final matching criterion. The adequacy of the propensity scores in controlling for treatment selection bias was demonstrated by testing for differences between matched patients in biological parameters likely to trigger RRT on reaching maximum RIFLE class.

The goodness of fit and the discrimination of the two logistic regression models used to derive a propensity score for RRT were evaluated by the Hosmer-Lemeshow (HL) test, and the c statistic (area under the receiver operating characteristics curve), respectively.

Results are expressed as numerical values and percentages for categorical variables, and as means and standard deviations (SD) or medians and quartiles [Q1-Q3] for continuous variables.

In the whole cohort, comparisons of patients with and those without RRT were based on chi-square tests for categorical data, and on Student's t-test or Wilcoxon's test for continuous data, as appropriate.

Comparisons between matched patients were based on univariate conditional logistic regression. Multivariate conditional logistic regression analysis was used to examine the association between RRT and subsequent hospital mortality, adjusting for variables potentially related to mortality that were not considered in the propensity regression (namely baseline characteristics that had a P value < 0.1 in univariate analysis, and the modified SOFA score (SOFA score - specific renal component) computed on the day maximum RIFLE class was reached).

Sensitivity analyses were performed to test whether any delay in RRT initiation could affect patients' prognosis. For that purpose, the timing of RRT was divided into three classes (less than 24 h, between 24 and 48 h, greater than 48 h after reaching maximum RIFLE class).

Since the use of the MDRD equation to estimate baseline creatinine values has not been validated in ICU patients, we also performed sensitivity analyses that included only patients with a normal serum creatinine value measured on ICU admission.

Wald χ2 tests were used to determine the significance of each variable. Adjusted odds ratios (ORs) and 95% confidence intervals (CIs) were calculated for each parameter estimate.

Analyses were computed using the SAS 9.1 software package (SAS Institute, Cary, NC, USA).

Over the study period, 10,911 patients with a single ICU stay were screened, of whom 2,272 were excluded for the following reasons: decision to withhold or withdraw life-sustaining treatments (n = 1,378, 12.6%), history of chronic kidney disease (n = 672, 6.2%), functional renal failure (n = 176, 1.6%), and RRT for extra-renal indications (n = 46, 0.4%).

Among the remaining 8,639 patients, 2,846 (32.9%) had AKI (1,025 (36%) R class patients, 830 (29.2%) I class patients, and 991 (34.8%) F class patients).

RRT was initiated in 545 (19.1%) AKI patients (41 (7.5%) R class patients, 110 (20.2%) I class patients, and 394 (72.3%) F class patients).

Patients who received RRT were younger, had higher severity scores, were more likely to be transferred from ward, and presented more comorbidities than patients who did not receive RRT (Table 1). Differences between patients with and without RRT according to the maximum RIFLE class reached during the ICU stay are shown in Additional files 1, 2, and 3.

AKI occurred early in the course of ICU stay. Three-quarters of the patients reached their maximum RIFLE within three days after ICU admission.

When a decision of RRT was made, RRT was started less than 48 h after reaching maximum RIFLE class in 479/545 (87.9%) patients. Continuous veno-venous hemofiltration/hemodiafiltration and intermittent hemodialysis were used as initial RRT modality in 345 (63.3%) patients and 200 (36.7%) patients, respectively.

Details on timings of AKI and RRT for each RIFLE class are shown in Tables 2 and 3.

Differences in parameters (measured on reaching maximum RIFLE class) likely to trigger RRT between patients who actually received RRT and those who did not are presented in Table 4. Patients with RRT had higher blood urea nitrogen, serum creatinine and kaliemia but their pH values were not significantly lower.

The two propensity models showed satisfying goodness of fit and discrimination (P values for the HL test: 0.39 and 0.52, c statistics: 0.80 and 0.78, in models 1 and 2, respectively). The percentage of matched patients was high despite numerous and strict matching criteria. In model 1, 383/545 (70%) patients who received RRT could be matched to 726 controls who did not receive RRT. In model 2, 376/545 (69%) RRT patients could be matched to 754 controls. In both models, there were no differences between patients with and those without RRT in biological parameters likely to trigger RRT on reaching maximum RIFLE class (Table 5), thus confirming the ability of the propensity scores to control for treatment selection bias. However, there remained differences in SAPS II, mSOFA, urine output and fluid accumulation that were thus adjusted for (Table 5).

RRT resulted in longer lengths of ICU stay after reaching maximum RIFLE class (see Additional files 4 and 5) but did not reduce mortality. Crude hospital mortality rates of patients with and without RRT were 45.1% and 23.4%, respectively (P < 0.001). Among patients who received RRT, 92 of the 338 survivors (27.2%) still needed renal support on ICU discharge.

After matching on the propensity scores, patients who received RRT still had higher mortality rates than their respective controls (model 1: 38.9% vs 22.2%, P < 0.001; model 2: 38% vs 18.3%, P < 0.001), in univariate analysis. After adjustment on confounding variables, RRT was not associated with a reduced hospital mortality, whatever its timing (Table 6). Additional files 6 and 7 show details according to the maximum RIFLE class reached during the ICU stay.

The sensitivity analyses that included only patients with a normal serum creatinine value measured on ICU admission yielded similar results as the full analysis (Additional file 8).