Study design
We conducted a single-center before-after cohort study with a within-subject design at the Hôpital Louis Pradel (Hospices Civils de Lyon, France). Consecutive patients were eligible over two periods: from 1st January to 31th December 2020 (pre-intervention period) and 15th February to 15th August 2021 (intervention period). All data were prospectively and automatically recorded via the IntelliSpace Critical Care and Anaesthesia software (V H.02.01, Philips Healthcare, Andover, MA, USA) or the patients’ computerized medical file (easily® V05.12.00.00, Hospices Civils de Lyon, France). Data were retrospectively collected during the pre-intervention period and prospectively collected during the intervention period. The two periods were separated by a run-in period of 6 weeks allowing to put the protocol in place. The study protocol was registered on clinicaltrials.gov (NCT05119361) and approved by the local ethics committee (IRB: 00013204). As the perfusion-based deresuscitation protocol was considered as a new standard of care, the institutional review board waived the need for written informed consent. According to French law, non-opposition to the use of patients’ health data for research purposes was systematically proposed and the data collection was approved by the Commission national de l’informatique et des libertés. The manuscript followed the STROBE guidelines (17) and the completed checklist is available in Additional file 1.
Participants and study protocol
Eligible patients had to fulfil the following inclusion criteria: age > 18, need for CRRT for acute kidney injury, cumulative fluid balance above 5% (based on weight-related cumulative fluid balance or computerized cumulative fluid balance as defined below), and norepinephrine equivalent dose < 0.5 µg/kg/min (18). Exclusion criteria were: pregnancy, active bleeding, chronic intermittent hemodialysis, stroke with coma, advanced directives to withhold or withdraw life-sustaining treatment, and patient’s opposition to the use of his/her personal health data. Patients meeting the inclusion criteria and admitted during the pre-intervention period were included in the control group, whereas those admitted during the intervention period were included in the early dry group. In the control group, the management of fluid balance during the deresuscitation phase (i.e. timing of initiation, intensity, monitoring, reason to discontinue fluid removal, etc.) was left to the clinician discretion since there was no institutional protocol. By contrast, the complete deresuscitation perfusion-based protocol of the early dry group is illustrated in Figure 1.
Outcomes
The primary outcome was the computerized cumulative fluid balance at day 5, at death, or at discharge, automatically calculated by the software and defined as the difference between total intake (cumulative volume of medication, enteral and parenteral feeding, fluid loading, and transfusion products) and total output (cumulative volume of diuresis, surgical drainage, and UFnet).
Secondary outcomes were: 1) weight-related cumulative fluid balance defined by weight variations with electronic bed weighing = (weight at day 5 or discharge - weight on the day of inclusion) / weight on the day of inclusion, 2) inclusion rate of patients included in the early dry group to estimate the population of eligible patients in a large RCT, defined as the ratio between included patients to admitted patients, 3) hemodynamic tolerance of the protocol defined both by the incidence of hypoperfusion in the early dry group (number of events per day) and by the maximal arterial lactate level and maximal norepinephrine dose until day 5 or discharge in the two groups, 4) exploratory outcomes: weaning of mechanical ventilation at day 7 among live patients, weaning of renal replacement therapy at day 30 among live patients, mortality rate at day 30, number of ventilator-free days at day 30, number of renal replacement therapy-free days at day 30, number of norepinephrine-free days at day 30, number of organ failure-free days at day 30 (see Additional file 2 for definitions).
Statistical analysis
Considering recruitment difficulties in a previous study using a comparable strategy (19), we did not determine a sample size a priori but rather a study period to determine the feasibility of a RCT (the potential rate of inclusion). Missing data concerning Sepsis Organ Failure Assessment (SOFA) score were handled as previously described (20): missing score components at baseline were assigned the score of 0, otherwise the last value was used until new data were available. To handle missing values at baseline of other variables, required for inverse probability weighted adjustment, we used a random forest based process to impute missing data. The maximum iteration was set to 10. No imputation was carried out for other variables recorded after baseline. Variables with a rate of missing values greater than 25% were not analyzed.
Data were expressed as mean (standard deviation, SD), median [interquartile range, IQR], or count (percentage), as appropriate. Characteristics of the two groups were compared using the Student t test or the Mann-Whitney U test for continuous variables, and a χ² test or the Fisher exact test for categorical variables. As the cohort was observational, we adjusted the data to the inverse probability of belonging to the control group or to the early dry group by using a propensity score built on a logit model. Variables initially included in the model were all pertinent variables with a P value less than 0.2. UFnet and delay between the admission and the inclusion were not included in the model as the difference observed is a direct consequence of the protocol application, nor was the acute respiratory distress syndrome due to the obvious collinearity with the COVID-19 status. Then we added age, arterial lactate, and vasoactive-inotropic score to balance groups regarding those three known prognostic factors (21–23). We considered no collinearity of variables if the square root of the variation to inflation ratio was less than 2. We then used the inverse probability to create a weighted pseudo population. To decrease the variability of the effect of the treatment, we truncated the weight to a maximum of the 99th percentile of all weights to decrease the impact of outliers. Each participant was weighted using the overlap weight approach, which down-weights individuals based on propensity score values. Covariate balance between the two groups was assessed after weighting, and we considered an absolute standardized difference of less than 0.1 as evidence of balance and 0.25 as acceptable balance considering the small sample size (24).
Statistical analyses were performed using R version 4.1.2 (R Core Team 2017, Vienna, Austria). IPW (25), missForest (26), and survey (27) packages were used. All tests were two-sided and a P value less than 0.05 was considered significant.