Conservative vs. Liberal Approach to Fluid Therapy of Septic Shock in Intensive Care

Fluid resuscitation is a key intervention in treatment of sepsis, but exact indications for fluid and amount of fluid administered is not established. Current guidelines for fluid therapy beyond 6 hours are vague and ungraded and most observational studies suggest harm with increasing positive fluid balance.

Objective To assess feasibility of a protocol comparing a conservative (trigger guided) vs. liberal (target guided) approach to fluid resuscitation in patients with septic shock after initial fluid resuscitation.

Design Multicentre, parallel group, centrally randomised, open label trial with adequate generation of allocation sequence, and adequate allocation concealment.

Sample size 150 included patients we will be needed to show a 1.7 L difference in fluid volumes between the groups based on the mean volume of resuscitation fluid given within first 5 days observed in the 6S trial of 5.3 L (SD 3.7 L) with a maximal type 1 and 2 error of 5% and 20% (power=80%), respectively.

Funding The trial is publicly funded by the Danish Council for Strategic Research

Statistical Analysis Plan for the Classic Trial

Outcome measures The outcome measure "Amount of resuscitation fluid given during ICU stay" has been changed from a secondary outcome measure to a co-primary outcome measure, which differs from the Classic Trial protocol. The Classic Trial intervention period is entire ICU stay and we consider it appropriate to have an outcome measure addressing this as co-primary outcome measure. This change has been approved by the Danish Ethical Committee and the Danish Health and Medicines Authorities (Applied for June 19 2015).Multiplicity issues will be addressed (see Analyses section).

Analyses

All statistical tests will be 2-tailed. Multiplicity adjustment. Dealing with multiplicity, the parallel gate keeping method with truncation parameter lambda = 0 will be used to adjust the observed (raw) P values for primary and secondary outcomes (Dmitrienko A, Tamhane AC, Bretz F. Multiple testing problems in pharmaceutical statistics. Chapman \& Hall/CRC biostatistics series (2010)).

By this approach the null hypotheses are divided into two families: F1 including null hypotheses related to the two co-primary outcomes and F2 including null hypotheses related to the secondary outcomes. The raw P values are then adjusted. If at least one of the adjusted P values in family 1 is less than the chosen level of significance the hypotheses in family 2 are also tested. If not the hypotheses in family 2 are all accepted without test. However, in all events all raw P values as well as the adjusted ones will be presented.

Lambda may be varied between 0 and 1. If the effect sizes of the primary outcomes (corresponding to the null hypotheses of F1) are uniformly high a lambda near 1 will help improve the overall power. On the other hand if the effect sizes are expected to vary across the endpoints, the overall power is likely to be maximized when lambda is small (Dmitrienko A, Tamhane AC, Bretz F. Multiple testing problems in pharmaceutical statistics. Chapman \& Hall/CRC biostatistics series (2010)).

We expect a degree of correlation between the two co-primary outcome measures somewhat in between full correlation and no correlation, so a conventional adjustment of the significance level (0.05/2=0.025) may result in a too conservative adjustment. Thus, we have chosen to adjust the level of significance by a factor in between a full Bonferroni adjustment and no adjustment at all, that is 0.05/1.5=0.033. In the above procedure the raw P values and not the significance level are adjusted and usually α (the significance level) is chosen to be 0.05. In family 1 the smaller raw P value is adjusted by multiplying it with 2. Therefore, we implement the above adjustment solving 2*0.0.033 ≤ level of significance => level of significance = 0.066 to secure that a raw P value ≤0.033 for a co-primary outcome will imply that the corresponding null hypothesis will be rejected.

Revised power calculation

The multiplicity adjustments for the co-primary outcome measures infer changes in the power calculations. The revised power calculations are based on 150 included patients with α=0.033 and β=0.80:

Outcome measure 1.1: Power to show a 1.8 L (opposed to 1.7 L with α=0.05) difference in fluid volumes between the groups based on the mean volume of resuscitation fluid given within first 5 days observed in the 6S trial of 5.3 L (SD 3.7 L) Outcome measure 1.2: Power to show a 4.1 L (opposed to 3.7 L with α=0.05) difference based on mean of 8.0 L (SD 8.1 L) total resuscitation fluid volume during ICU stay days in the 6S trial.

We regard the revised power to be sufficient to address the research question; thus, the sample size will not be changed.

1. Analysis of outcome measures

Two analyses will be done for the co-primary outcome measures (1,2):

1. an analysis adjusted by the stratification variable (site) - primary analysis
2. an analysis adjusted by the stratification variable and baseline covariates ((a) surgery during current hospitalisation but prior to randomisation Y/N, (b) Age, (c) more than 5 L of fluid (crystalloids, colloids and blood products combined) given in the 24 hours prior to randomisation Y/N, (d) highest dose of noradrenalin in the 24 hours prior to randomization, (e) estimated weight at randomisation For exploratory outcome (10) we will perform both an unadjusted analysis (for log rank test) and an analysis adjusted by the stratification variable site.

The remaining outcome measures will only be analysed adjusted by the stratification variable (site).

Co-primary outcomes (1,2), secondary outcomes (3-6) and exploratory outcome (15) will be analysed using the general linear model.

The exploratory outcomes (9) and (13) will be analysed using logistic regression.

The exploratory outcome (10) will be analysed using Kaplan Meier survival plots and the log rank test. Adjusted analysis will be done using Cox regression model stratified by site.

Secondary outcome (7) will not be compared between intervention groups , because major protocol violations can only occur in the conservative (Trigger-guided) group.

Secondary outcomes (8), and exploratory outcomes (11), (12) and (14) will be analysed using the Poisson distribution with link = log and offset or the negative binomial distribution with link=log and offset as appropriate. As a sensitivity analysis the two groups will also be compared using a non-parametric test (van Elteren test adjusted for site) and major differences in the results obtained by the two approaches will be discussed.

1. Sensitivity analyses

The primary outcomes will be analyzed using each of the two per-protocol populations.

Populations Intention-to-treat population: All randomised patients except those who withdraw their consent for the use of data.

Per-protocol population:

All randomised patients except patients having one or more protocol violations defined as:

1. One or more resuscitation fluid boluses given without fulfilment of one or more of the Classic-criteria in the Conservative (Trigger-guided) group. OR
2. Use of colloids (either Albumin or synthetic colloids) for resuscitation OR
3. Monitoring revealed that one or more in- or exclusion criteria were violated OR
4. Stopped/withdrawn patients

Subgroups:

1. Patients with more than 5 L of fluid (crystalloids, colloids and blood products combined) given in the 24 hours prior to randomisation

The results of the subgroup analysis will be presented if P of test of interaction between subgroup indicator and intervention group indicator for primary outcome is < 0.05. The P-value of the test of interaction will be presented regardless.

Missing Data

Missing primary outcome data:

We do not expect missing data on the co-primary outcome measures (1,2). Only complete case analysis will be made.

Missing secondary outcome data We do not expect missing data on the secondary outcome measures 7 and 8. Only complete case analysis will be made.

Missing data on secondary outcomes 3-6: Since the predictors (centre indicator and intervention indicator) will not be missing only the outcome may be missing. In this case a complete case analysis will be unbiased since the cases with outcome missing carry no information. However, auxiliary variables (i.e. variables not included in the analytical model such as e.g. other outcomes) may be correlated with the outcome and their inclusion in the analysis will improve the efficiency. This possibility is best dealt with using a structural equation model for the regression analysis with direct maximum likelihood estimation and inclusion of the auxiliary variables (the SAS proc calis for continuous dependent variable may be used). However, the data may still be missing not at random. Therefore, a sensitivity analysis estimating the range of potential bias that may be caused by data missing not at random is done where the missing values in one group are replaced by the minimum value in the whole material and the missing values in the other group are replaced by the maximum value in the whole material and vice versa. The corresponding P values will be estimated. The standard error of each of the two estimates of the regression coefficient will be replaced by the corresponding standard error from the complete case analysis (or the direct ML analysis if auxiliary variables are used) if it is smaller than the former

Missing baseline data

Fluids given prior to randomisation Yes/no Some patients may have missing data on fluids given prior to randomisation. In this case it is a regression of each of the co-primary outcomes on centre, and the above mentioned baseline covariates of which only fluids given prior to randomization Yes/no has missing values. As long as the probability of missing data on the predictor is independent of the outcome a complete case analysis will give unbiased results even if the probability depends on the missing predictor values (i.e., data are NMAR) (Allison PD Missing data Sage publications (2001)). Therefore, if the mean values of the outcome do not differ significantly (P < 0.10) between those patients with missing values and those without missing values a complete case analysis will be done. If not multiple imputation of the missing baseline variable will be done using monotone logistic regression.