Net ultrafiltration intensity and mortality in critically ill patients with fluid overload

We conducted a retrospective study using a large tertiary care academic medical center ICU database: the High-Density Intensive Care dataset, details of which have been published elsewhere (Additional file 1: S1) [1, 17, 18]. The study population included adults admitted to medical, cardiac, abdominal transplant, cardiothoracic, surgical, neurovascular, neurotrauma and trauma ICUs during July 2000 through October 2008. We included patients with AKI receiving RRT who had a cumulative fluid balance ≥ 5% prior to RRT initiation (Additional file 1: Figure S1). We extracted the daily fluid balance before and for the duration of RRT (Additional file 1: S2), hourly mean arterial pressure (MAP) and vasopressor type and dose (Additional file 1: S3) during RRT. The University of Pittsburgh’s institutional review board approved the study.

For patients receiving continuous renal replacement therapy (CRRT), we first extracted data on the total duration (in hours) of any form of CRRT (i.e., continuous venovenous hemodiafiltration (CVVHDF), continuous venovenous hemofiltration (CVVH), continuous venovenous hemodialysis (CVVHD) and slow continuous ultrafiltration (SCUF)). We then determined the UF volume produced and the amount of substitution fluids given each hour for patients receiving CVVHDF and CVVH. The UF^NET each hour was calculated as the difference between the UF volume and the volume of substitution fluids [19]. For patients receiving CVVHD and SCUF, UF^NET corresponded to the UF volume removed. We then calculated the total number of days of CRRT for each patient based on the hourly duration of CRRT and the total UF^NET.

For patients receiving intermittent hemodialysis (IHD), we extracted the total number of IHD sessions and the UF volume removed per session from the time of ICU admission to the end of ICU stay. We excluded patients if they received IHD prior to ICU admission. UF^NET corresponded to the volume ultrafiltered during each session. We then expressed the total number of IHD sessions as the number of days for each patient. Subsequently, we estimated the UF^NET intensity using the equation:

For instance, if an 80-kg patient is on CVVH with an UF rate of 2000 ml/h and substitution fluid of 1500 ml/h, the total UF^NET produced is 500 ml/h (2000 – 1500 = 500 ml) or 500 × 24 = 12,000 ml/day. The total UF^NET produced for 5 days is 12,000 × 5 = 60,000 ml. Thus, the total UF^NET intensity is [60,000 / (80 × 5)] = 150 ml/kg/day. During CVVHD and IHD, the UF volume is equivalent to UF^NET.

The primary outcome was 1-year mortality from the index ICU admission and mortality data were obtained from the Social Security Death Master File [20]. We chose 1-year mortality because our prior work showed that a positive fluid balance was associated with risk of death at 1 year and use of renal replacement therapy was associated with lower risk of death in patients with a positive fluid balance [1]. Secondary outcomes included hospital length of stay, hospital mortality and renal recovery. Renal recovery was defined as alive and independent from RRT at 1 year. Dialysis dependence data were obtained from the US Renal Data System [21].

We stratified UF^NET intensity into three groups because of the nonlinear (i.e., J-shaped) association between UF^NET intensity and hospital mortality (Additional file 1: Figure S2). We defined UF^NET ≤ 20 ml/kg/day as “low” intensity, UF^NET > 20 to ≤ 25 ml/kg/day as “moderate” intensity and UF^NET > 25 ml/kg/day as “high” intensity. Categorical variables were compared using the chi-squared test, and continuous variables using-one way analysis of variance and the Kruskal–Wallis test. We assessed time-to-mortality censored at 1 year using Kaplan–Meier failure plots.

We used three methods to examine the association between UF^NET intensity and mortality. First, we fitted logistic regression and estimated risk-adjusted odds ratios (AORs) for high and moderate intensity, compared with low intensity UF^NET (reference), on 1-year mortality. Second, we fitted Gray’s survival model [22, 23] to estimate risk-adjusted hazard ratios (AHRs) for time to mortality using four time nodes and five intervals (Additional file 1: S4). We adjusted for differences in age, sex, race, body mass index, history of liver disease and sequela from liver disease, admission for liver transplantation, admission for surgery, baseline glomerular filtration rate, Acute Physiologic and Chronic Health Evaluation (APACHE) III score, presence of sepsis, use of mechanical ventilation, percentage of FO before initiation of RRT, oliguria before initiation of RRT, time to initiation of RRT from ICU admission, MAP on first day of RRT initiation, cumulative vasopressor dose and cumulative fluid balance during RRT, first RRT modality and duration of RRT.

Third, in order to account for indication bias, we conducted a propensity score-matched analysis. Since the mortality associated with moderate (> 20 to ≤ 25 ml/kg/day) vs high (> 25 ml/kg/day) or moderate (> 20 to ≤ 25 ml/kg/day) vs low (≤ 20 ml/kg/day) intensity UF^NET was not different (Table 1), we combined the moderate and low-intensity groups into a single low-intensity group (reference). We then matched the low-intensity UF^NET (≤ 25 ml/kg/day) with the high-intensity UF^NET (> 25 ml/kg/day) using propensity scores on a 1:1 basis without replacement, creating 258 matched pairs (Additional file 1: S5).

We performed five sensitivity analyses and two subgroup analyses. First, we restricted the UF^NET intensity only up to 72 h from initiation of RRT. Second, we used an alternative definition of UF^NET intensity moving the threshold down as follows: low, < 15 ml/kg/day; moderate, 15–20 ml/kg/day; and high, > 20 ml/kg/day. Third, we moved the threshold up: low, < 25 ml/kg/day; moderate, 25–30 ml/kg/day; and high, > 30 ml/kg/day. Fourth, we divided the cohort into tertiles: low, ≤ 16.7 ml/kg/day; moderate, 16.7 to ≤ 27.7 ml/kg/day; and high, > 27.7 ml/kg/day. Fifth, we performed quantitative bias analysis to assess the magnitude of a hypothetical unmeasured confounder that would be necessary to account for the association between UF^NET intensity and risk-adjusted mortality (Additional file 1: S6) [24, 25].

Sixth, we restricted our analyses only to the subgroup of patients with > 20% FO. Seventh, we confined our analysis of UF^NET intensity to the hour (i.e., ml/kg/h) instead of the day among the subgroup of patients who only received CRRT as follows: low, < 0.5 ml/kg/h; moderate, 0.5–1.0 ml/kg/h; and high, > 1 ml/kg/h. Statistical analyses were performed using SAS 9.3 (SAS Institute, Cary, NC, USA), Gray’s model used R 3.2.1, and quantitative bias analysis was performed using STATA 15 (STATCorp., TX, USA). All hypotheses tests were two-sided with a significance level of p < 0.05.

Of 45,568 patients, we excluded patients with no available baseline weight (n = 2214), ICU duration ≤ 48 h (n = 18,032), death within 72 h of ICU admission (n = 663), chronic dialysis (n = 2386), admission for or with history of renal transplantation (n = 1232), serum creatinine ≥ 3.5 mg/dl within 1 year of hospitalization (n = 147) and missing data on fluid balance (n = 2810). Of 18,084 patients in whom cumulative fluid balance data were available, we excluded those with cumulative fluid balance < 5% of body weight (n = 9900). Of patients with cumulative balance ≥ 5% of body weight (n = 8184), we excluded those who did not receive RRT (n = 7023) and patients without data on UF^NET (n = 86) (Additional file 1: Figure S1).

Of 1075 patients, the distribution of low, moderate and high-intensity UF^NET groups was 44.2%, 15.2% and 40.4%, respectively. Minor differences were noted among male sex, race and body mass index between the groups (Table 1). There was a higher prevalence of liver disease (34.5%), sequela from liver disease (28.8%) and liver transplantation (21.5%) among those with low-intensity UF^NET. There was a higher prevalence of oliguria in those who received moderate and high-intensity UF^NET. Patients in the low-intensity UF^NET group had lower MAP compared with the moderate and high-intensity UF^NET groups (Table 1 and Additional file 1: Table S1).

Cumulative FO before RRT initiation was lowest in the low-intensity group, compared with the moderate and high-intensity UF^NET groups (15.6% vs 17.3% vs 21% of body weight, respectively, p < 0.001; Table 2). Following initiation of RRT, the median cumulative FB for the low, medium and high-intensity UF^NET groups was 13.5 vs 22 vs 19 l, p < 0.001; Table 2). During RRT, the MAP was lower and the cumulative vasopressor dose was higher in the low-intensity UF^NET group compared with the moderate and high-intensity UF^NET groups (Table 2 and Additional file 1: Table S1).

The median duration of RRT for the low, moderate and high-intensity UF^NET groups was 4.7 vs 8.7 vs 7 days, respectively (p < 0.001). The median duration of CRRT was 3.9 vs 5.8 vs 5.9 days (p < 0.001) and the median UF^NET volume was 3.4 vs 11.6 vs 16.2 L (p < 0.001). The median duration of IHD was 2 vs 7 vs 4 days (p = 0.004) and the median UF^NET volume was 5.5 vs 12.6 vs 9.2 L (p < 0.001). The median duration of RRT for patients who received both CRRT and IHD was 14.7 vs 15.2 vs 10.7 days (p < 0.001) and the median UF^NET volume was 19.5 vs 27.9 vs 26.6 L (p < 0.001). The median hospital length of stay was 32 vs 37.5 vs 37 days (p < 0.001) (Table 2). This shorter length of stay among patients with low-intensity UF^NET was primarily due to higher mortality in this group. However, there was no difference in renal recovery at 1 year (25.1% vs 28.9% vs 31.8%, p = 0.078) as well as within the subgroup of survivors at 1 year (82.6% vs 72.7% vs 78.4%, p = 0.25) between the three groups.

The crude hospital and 1-year mortality was higher among the low-intensity group compared with the moderate and high-intensity UF^NET groups: 69.7% vs 60.2% vs 59.4% (p = 0.003), respectively (Table 2, Fig. 1a). Using logistic regression, high-intensity compared with low-intensity UF^NET was associated with lower 1-year mortality (AOR 0.61, 95% CI 0.41–0.93, p = 0.02, C-statistic 0.811; Table 3 and Additional file 1: Table S2). This association persisted using UF^NET as a continuous variable (AOR 0.98, 95% CI 0.97–0.99, p = 0.005; Additional file 1: Table S3). Compared with UF^NET of 0–5 ml/kg/day, increasing UF^NET intensity was associated with a trend toward lower odds of death (C-statistic – 0.813; Fig. 1b), whereas moderate-intensity compared with low-intensity UF^NET was not associated with mortality (AOR 0.81, 95% CI 0.48–1.35, p = 0.41; Additional file 1: Table S2).

Using Gray’s model, high-intensity compared with low-intensity UF^NET had variable association with mortality. Early on after ICU admission, high-intensity UF^NET was associated with lowest risk of death that was subsequently attenuated over time, but nevertheless persisted up to 39 days after ICU admission (AHR range 0.50–0.73, p < 0.001; Table 4 and Additional file 1: Figure S3A). Subsequently, between 39 and 365 days, high-intensity UF^NET was not associated with mortality (AHR range 0.76–1.02). High-intensity compared with moderate-intensity UF^NET was only associated with lower risk of death up to 15 days (AHR 0.53, 95% CI 0.33–0.86; Additional file 1: Table S5 and Figure S3C).

After propensity matching, 258 matched pairs were created wherein patients with UF^NET intensity ≤ 25 ml/kg/day had similar baseline characteristics compared with UF^NET intensity > 25 ml/kg/day, except for cumulative vasopressor dose (Additional file 1: Table S4). Patients with UF^NET intensity > 25 ml/kg/day compared with ≤ 25 ml/kg/day had lower 1-year mortality (57% vs 67.8%, p = 0.01; Fig. 2), which persisted after adjusting for vasopressor dose (AOR 0.63, 95% CI 0.44–0.90, p = 0.011).

When UF^NET intensity calculation was limited within 72 h of initiation of RRT, high-intensity UF^NET was associated with lower mortality (AOR 0.56, 95% CI 0.35–0.88, p = 0.013; Table 5). Using the alternative thresholds of low, moderate and high-intensity UF^NET of < 15 ml/kg/h, 15–20 ml/kg/h and > 20 ml/kg/h, respectively, we found UF^NET intensity > 20 ml/kg/h was associated with lower mortality (AOR 0.63, 95% CI 0.41–0.97, p = 0.038). Similar results were found moving the threshold up (AOR 0.58, 95% CI 0.34–0.99, p = 0.04; Table 5) and using tertile cutoff values (AOR 0.61, 95% CI 0.39–0.97, p = 0.037; Table 5). The quantitative bias analysis indicated that our results would be robust unless an unmeasured confounder was at least twice as prevalent among patients who received high-intensity UF^NET as among those with low-intensity UF^NET (Additional file 1: Figure S4). The unmeasured confounder should have an OR < 0.7 (i.e., reduced risk of death by more than 30%) to mask a null association between high-intensity UF^NET and risk-adjusted mortality (Additional file 1: S6 and Figure S4).

High-intensity UF^NET was associated with a trend toward lower mortality among patients with > 20% FO (AOR 0.52, 95% CI 0.26–1.05, p = 0.07; Table 5). For patients receiving CRRT, UF^NET intensity > 1.0 ml/kg/h compared with UF^NET intensity < 0.5 ml/kg/h was associated with lower odds of death (AOR 0.41, 95% CI 0.24–0.71, p = 0.0013).