Professional Documents
Culture Documents
Affiliations:: AJRCCM Articles in Press. Published March 29, 2022 As 10.1164/rccm.202111-2484OC
Affiliations:: AJRCCM Articles in Press. Published March 29, 2022 As 10.1164/rccm.202111-2484OC
(BaSICS) Study
Fernando G. Zampieri, MD, PhD,1,2 Flávia R. Machado, MD, PhD,2,3 Rodrigo S. Biondi, MD,2,4
Flávio G. R. Freitas, MD,2,5 PhD, Viviane C. Veiga, MD,2,6 PhD, Rodrigo C. Figueiredo, MD,7
Wilson J. Lovato, MD,8 Cristina P. Amêndola, MD, PhD,9 Ary Serpa-Neto, MD, PhD,2,10 Jorge L. R.
Paranhos, MD,11 Eraldo A. Lúcio, MD,PhD,12 Lúcio C. Oliveira-Júnior, MD,13 Thiago C. Lisboa, MD,
PhD,2,14 Fábio H. Lacerda, MD,15 Israel S. Maia, MD,1,2,16 Cintia M. C. Grion, MD, PhD,2,17 Murillo
S. C. Assunção, MD, PhD,10 Airton L. O. Manoel, MD, PhD,18 Thiago D. Corrêa, MD, PhD,2,10
Marco Antonio VA Guedes, MD,19 Luciano C. P. Azevedo, MD, PhD,2,20 Tamiris A Miranda, MSc1,
Lucas P. Damiani, MSc,1 Nilton Brandão da Silva, MD, PhD,21 Alexandre B. Cavalcanti, MD, PhD1,2
Affiliations:
3 Department of Anesthesiology, Pain and Intensive Care, Universidade Federal de São Paulo
7 Hospital Maternidade São José, Centro Universitário do Espírito Santo (UNESC), Colatina,
Brazil.
8 Hospital das Clínicas, Faculdade de Medicina de Ribeirão Preto, Universidade de São Paulo,
11 Santa Casa de Misericórdia de São João Del Rei, São João Del Rei, Brazil.
12 Hospital São Francisco, Santa Casa de Porto Alegre, Porto Alegre, Brazil.
14 Hospital Santa Rita, Santa Casa de Porto Alegre, Porto Alegre, Brazil.
Londrina, Brazil.
Corresponding author: Fernando G Zampieri, MD, PhD, hcor Research Institute, Rua Abilio Soares
Funding/Support: This trial was supported by Brazilian Ministry of Health through the “Programa de
were supplied by Baxter Hospitalar. A.B.C. received a research scholarship from CNPq – Brazil
Role of the Funder/Sponsor: Baxter Hospitalar had no role in the design and conduct of the study.
The steering committee was responsible for the design and conduct of the study; collection,
management, analysis, and interpretation of the data; preparation, review, or approval of the
manuscript; and the decision to submit the manuscript for publication. The Brazilian Ministry of
Health approved the study, but was not involved in design, analysis, interpretation, or decision to
Conflicts of Interest: LCPA reports consulting fees from Halex-Istar, lecture fees from Pfizer and
Baxter and grants for research investigator from Brazilian Ministry of Health and Aché
Pharmaceuticals, both unrelated to this work. No other conflicts of interest were reported.
Author Contributions: Drs Zampieri and Cavalcanti had full access to the data in the study and take
responsibility for the integrity of the data and the accuracy of the data analysis.
Concept and Design: Zampieri, Cavalcanti, Damiani, Corrêa, Azevedo, Machado, Assunção,
Brandão.
Critical revision of the manuscript for important intellectual content: All authors
This article has an online data supplement, which is accessible from this issue's table of
At A Glance
Scientific Knowledge on the Subject: Balanced solutions may be associated with better outcomes in
critically ill patients. It has been suggested that potential benefits of balanced solutions may be more
significant if they are used before ICU admission and are maintained as preferred fluid during ICU
stay. Randomized controlled trials that allocated patients to receive either 0.9% saline or balanced
What This Study Adds to the Field: In this secondary analysis of the BaSICS trial, we found that
type of fluid patients received for initial resuscitation before enrollment appeared to mediate potential
benefits of balanced solution use in the ICU. There was a high probability that balanced solution use
was associated with lower 90-day mortality in patients that exclusively received balanced solutions
before study enrollment. This benefit was more apparent in patients with unplanned admission due to
Abstract
Rationale: The effects of balanced crystalloid vs saline on clinical outcomes for intensive care unit
patients may be modified by the type of fluid patients received for initial resuscitation and by the type
of admission.
Objectives: To assess whether results of a randomized controlled trial could be affected by fluid use
Methods: Secondary post-hoc analysis of the Balanced Solution in Intensive Care (BaSICS) trial,
which compared a balanced solution (Plasma-Lyte 148) to 0.9% saline in intensive care unit. Patients
were categorized according to fluid use in the 24 hours before enrollment in four groups: balanced
solutions only; 0.9% saline only; mix of both, and no fluid before enrollment, and according to
admission type (planned, unplanned with sepsis and unplanned without sepsis). The association
between 90-day mortality and the randomization group was assessed using an hierarchical logistic
Bayesian model.
Measurements and Main Results: 10,520 patients were included. There was a low probability that
the balanced solution was associated with improved 90-day mortality in the whole trial population
(odds ratio, 0.95; 89% credible intervals [CrI], 0.66-1.51; probability of benefit, 0.58); however,
probability of benefit was high for patients that received only balanced solutions before enrollment
(regardless of admission type, odds ratio, 0.78, 89% CrI, 0.56,1.03; probability of benefit, 0.92),
mostly due to a benefit in unplanned admissions due to sepsis (odd ratio, 0.70; 89% CrI, 0.50-0.97;
probability of benefit, 0.96) and planned admissions (odds ratio, 0.79; 89% CrI, 0.65-0.97; 0.97
Conclusion: There is a high probability that balanced solution use in the ICU reduces 90-day
mortality in patients that exclusively received balanced fluids before trial enrollment.
Introduction
care unit patients remains uncertain . In the Balanced Solutions in Intensive Care (BaSICS) trial, ICU
patients were randomized to receive either 0.9% saline or Plasma-Lyte 148® (balanced solution) as
preferred fluid for maintenance, resuscitation, and dilutions during ICU stay, but trial protocol
controlled neither the volume nor the type of fluid administered prior to enrollment, which occurred
after intensive care unit admission. [1]. This trial rendered neutral results for its primary endpoint (90-
day survival). In contrast, another large cluster randomized trial suggested a possible benefit of
balanced solutions using a composite endpoint [3], which was more pronounced in the subgroup of
septic patients, specially if balanced solutions were used before ICU admission [3,4]. It is therefore
conceivable that the effect of balanced solutions could be moderated by admission type and fluid use
before enrollment.
In this secondary post-hoc analysis of the BaSICS trial, we explored the association between
fluid use before enrollment, admission type and the effect of balanced solution versus 0.9% saline in
critically ill patients. We hypothesized balanced solutions would be associated with lower mortality at
90-days in the subgroup of patients that received only balanced solutions before enrollment and that
these effects would be different according to admission types; in particular, we hypothesized that
potential benefits of balanced solutions would be greater in the subgroup of septic patients who
previously received only balanced solutions [3,4]; these effects could be related to chloride toxicity or
Methods
Study Design: Post-hoc secondary analysis of a multicenter randomized clinical trial comparing a
balanced solution (Plasma-Lyte 148) to 0.9% saline in critically ill patients (BaSICS Trial).
Population: BaSICS included patients admitted in the intensive care units that required at least one
fluid expansion; who were not expected to be discharged the next day, and that had at least one
additional risk factor (age above 65 years; hypotension; presence of sepsis; need for mechanical
presentation; or a diagnosis of liver cirrhosis or acute liver failure). Detailed information on inclusion
Fluid use in the 24 hours before enrollment was collected in the case report form as two
sequential questions: Whether there was any prescription of fluid therapy (not for dilution or
maintenance) in the past 24 hours on medical records (electronic of physical charts, transfer notes,
surgical records, etc), and, if yes, the volume of fluid of both 0.9% saline and balanced solutions
(defined as Lactated Ringer and/or Plasma-Lyte 148) used. These variables were obligatory in the
case report form and were checked for completion during site monitoring. Patients were then
categorized according to intravenous crystalloid fluid use in the 24 hours before enrollment in four
groups: (1) Patients that did not receive any saline and received exclusively balanced solutions, in any
volume; (2) Patients that received exclusively 0.9% saline before enrollment, in any volume; (3)
Patients that received a mix of balanced solutions and 0.9% saline (“Mixed Fluid”), and (4) patients
that did not receive intravenous crystalloid fluid before enrollment (“No fluid”). Admission type was
categorized as unplanned admission due to sepsis (defined as suspicion of infection plus organ
Interventions: Patients enrolled in BaSICS were randomized to receive 0.9% saline or balanced
solution as fluid of choice for all fluid challenges, maintenance, and drug dilutions (above 100 mL)
during ICU stay up to 90 days after enrollment. Physicians, patients, and individuals who assessed the
outcomes were blinded to the assigned treatment. Patient management, including the decision to
perform fluid challenges, was left to the discretion of the attending physician in the trial. Fluid use
before enrollment was noted in the case report form as volume of either 0.9% saline or balanced
solution (including Lactated Ringer or Plasma-Lyte 148) in the 24 hours before enrollment. See [2]
Endpoints: The primary endpoint for this secondary analysis was 90-day mortality. The secondary
endpoint was days alive free of kidney replacement therapy (KRT) up to 28 days after randomization.
Statistical analysis:
Primary endpoint: Univariate analyses are presented as table and exploratory figures. The association
between the primary endpoint and the randomization group was assessed using a hierarchical logistic
Bayesian model adjusted by group of fluid use before enrollment, admission type, the intervention,
and enrolling site as random intercept. An interaction between fluid use before enrollment, admission
type, and intervention arm was added. Three sets of priors were applied for the log odds ratio of the
intervention [6], all assuming a normal distribution of the log odds ratio [log(OR)]: one moderate
strength skeptical prior (mean 0 and standard deviation of 0.355), one moderate strength optimistic
prior (mean -0.182, standard deviation of 0.175; compatible with an odds ratio of 1/1.20 for benefit
and allowing 0.15 probability of harm), and one moderate strength pessimistic prior (mean 0.182,
standard deviation of 0.175; compatible with an odds ratio of harm of 1.2 and allowing 0.15
probability of benefit). Reasons for these priors are described on Table E1. Other priors, including
priors for interactions, were set as normal mean zero with standard deviation of one. Due to the
presence of interactions, results were obtained by sampling 4,000 conditional posterior probabilities
and providing the following metrics in different possible combination scenarios: (1) The median odds
ratio with 89% credible interval; (2) Probability of benefit of the intervention (P(OR) < 1); (3)
Probability odds ratio ranging from 1/1.1 to 1.1 (1/1.1<P(OR)<1.1); (4) Probability odds ratio below
1/1.25 (P(OR)<1/1.25); and (5) Absolute differences of predicted probabilities for possible scenarios.
We also display the 95% credible intervals for the primary endpoint in the main analysis in the tables,
Secondary analyses: A secondary analysis was performed according to the admission type and
evaluated the effect of total volume of fluid used in the 24 hours before enrollment, the percentage of
fluid infused as 0.9% saline, and the randomization arm. This analysis was designed to assess whether
a “contamination” effect occurred, that is, whether a continuous assessment of the percentage of fluid
used as saline could moderate the effect of randomization arm on the primary endpoint. Details are
shown in the online data supplement. Results are reported graphically as the conditional predicted
probabilities in potential scenarios (volume of fluid used and percentage given as 0.9% saline before
enrollment).
Sensitivity Analysis: One sensitivity analysis for the primary endpoint was performed after excluding
patients with traumatic brain injury, a population that may have been harmed by balanced solutions in
the main trial [2]; this analysis was performed for both the primary main analysis and the secondary
(continuous analysis). We also performed a sensitivity analysis based on frequentist methods (details
are shown in online data supplement) and another based on flat priors for all predictors for the
primary endpoint.
Secondary endpoint: Days alive and free of kidney replacement therapy up to 28 days were assessed
using a Bayesian beta binomial model with the same adjustment as for the main model. Methods are
Simulation Analysis: Due to the use of a complex three-way interaction model for the primary
benefit occurring under an absence of effect of balanced solutions after enrollment for the primary
endpoint (“type 1 events”). Details are shown in the Online Data Supplement.
Missing value policies: We used the same data set used for the main trial analysis, which included
imputed values for missing primary outcome for eleven patients, and admission type imputation for
17 patients, as described in [2]. Seventy-two (<1%) patients had unknown information on KRT; we
imputed this as absence of use of KRT. There were no missing values on whether fluid was used
before enrollment.
All analyses were performed using R version 4.1.1 [7] using packages brms [8] and tidybayes [9].
Code for the primary analysis and simulation are available in the online data supplement.
Results
A total of 10,520 patients were included in the analysis. Of all included patients, 3,202
received only balanced fluids before enrollment, 2,096 received only saline, 1,862 received a mix of
balanced solutions and saline and 3,360 did not receive crystalloids. Overall patient features according
to subgroups defined by fluid used before enrollment and randomization arm are described in Table 1
(aggregated values stratified only according to fluid use before enrollment is shown in Table E2). A
box plot of fluid use according to admission type and fluid use group in Figure E1. There was a weak
association between volume of 0.9% saline before enrollment and baseline chloride levels (Figure
E2). Trends in serum chloride levels for patient who had their serum chloride values measured are
shown in Figures E3-5 and percentage of patients with measured chloride that developed
hyperchloremia (defined as serum chloride above 110 mEq/L) is shown in Figures E6-8. Patients that
exclusively received balanced solutions before enrollment or that received a mix of balanced solutions
and 0.9% saline were more frequently admitted after elective surgeries, less frequently had sepsis, had
lower illness severity scores, and had lower mortality (Table E2). In univariate analysis, patients that
received balanced solutions or a mix of balanced solutions and 0.9% saline before enrollment had a
numerically lower mortality than those who received no fluids or exclusively 0.9% regardless of
Results for the primary endpoint is shown in Table 2, Figures 1 and 2, and Figures E10-15.
The results of the Bayesian reanalysis considering the skeptical neutral prior are shown in Table 2
(results of both optimistic, pessimistic, and flat priors are shown in Tables E2, E3, and E4 in the
Online Data Supplement, respectively). Results are presented in Table 2 for all patients and according
to all possible combinations of admission type and fluid use before enrollment. The probability that
being randomized to balanced solutions was associated with improved outcome in the whole
population was 0.58 (difference in mortality 0, 89% CrI -0.08 to 0.09; odds ratio 0.95, 89% CrI 0.66
to 1.51). These results are also shown separately according to fluid use before enrollment and
admission type on Figures E10 and Figure E11, respectively, and then sequentially according to
combinations of groups of fluid use before enrollment and admission type (Figures E12-E15).
Figure 1 shows the odds ratio and their 89% and 95% credible intervals for all possible
scenarios arising from the hierarchical Bayesian model shown in Table 2. The conditional
distributions of the odds ratio for balanced solutions versus saline in the 12 possible combinations are
shown in Figure 2. For patients that received only balanced solutions before enrollment, the
probability of benefit regarding mortality was high (0.92; Table 2, Figure E10 panel A), being very
high in planned admissions (0.97) and in unplanned ICU admission with sepsis (0.96), and less
pronounced in patients with unplanned admissions not due to sepsis (0.84) (Figure E11). . In no
admission type the overall probability that balanced solutions were associated with reduction in 90-
day mortality was above 0.90 (Table 2; Figure E12) In a sensitivity analysis excluding patients with
traumatic brain injury, there was an increase overall probability of benefit of balanced solutions in the
trial (0.58 to 0.63), driven by an increase in the probability of benefit in patients with unplanned
admissions without sepsis that only received balanced solutions before enrollment (0.84 to 0.94; Table
E4).Results were slightly affected by different priors (Table E2 and Table E3). For the optimistic
prior, all results were similar with no difference in interpretation of results. The pessimistic prior also
yielded similar results, with the exception of the posterior for balanced solutions on elective
admissions that received exclusively balanced solutions before ICU admission where a reduction in
probability of benefit was observed (0.97 to 0.85, Table E3). The flat prior, as expected, provided less
Results for the continuous analysis is shown in Figure 3 and in the online data supplement
(Figures E16-18). Overall, conditional predicted mortality was lower for balanced solutions in
patients that received a lower percentage of fluid before enrollment as 0.9% saline. This effect was
apparent across conditional total volumes of fluid used, being more evident for larger volume in
planned admissions. For unplanned admissions without sepsis, an invert association occurred for
higher volumes of fluid infused before enrollment, with higher mortality for balanced solutions; this
trend was reduced after excluding traumatic brain injury patients (which were mostly coded as
unplanned admissions without sepsis, shown in online data supplement Figure E19).
Frequentist analysis results are shown in the online data supplement (Tables E5 and E6, and
Figure E20); The only scenario were p values for the association between being randomized to receive
balanced solution versus 0.9% saline was <0.05 was on the subgroup of patients with sepsis that only
received balanced solutions before enrollment (OR 0.613; 95% confidence interval 0.330-0.895).
Average effects under frequentist framework are similar to Bayesian analysis with flat priors, being
both less conservative than the neutral prior used on the main analysis.
Results of days alive and free of KRT at 28 days are shown in Table E7. In no scenario the
probability that balanced solutions were associated with at least one day more alive and free of KRT
at 28 days was above 0.90. The higher probability of benefit found occurred in patients with
unplanned admission due to sepsis, where the probability of having one day more alive and free of
Results of the simulations for assessing potential frequency of type 1 events are shown in the
online data supplement. Probability of obtaining results as extreme as those found under the
assumption of absence of effect for the (1) all patients that received balanced solutions before
enrollment, (2) patients with unplanned admission due to sepsis that only received balanced solutions
before enrollment, and (3) planned admissions that only received balanced solutions before
enrollment were 0.065, 0.003, and 0.028, respectively. Details are shown on online data supplement.
Discussion
solution versus 0.9% saline in patients admitted to the intensive care unit, overall benefit of balanced
solutions appeared to be moderated by fluid use before enrollment and admission type. The overall
probability of benefit in the whole population was unremarkable, compatible with the original
publication [2]. However, important differences regarding the effect of the intervention according to
fluid use before enrollment and admission type were observed. There was a high probability of benefit
in patients randomized to balanced solutions in the subgroup of patients that received exclusively
balanced solutions before enrollment, independently of admission type (0.92); this benefit appeared to
be mostly driven by a reduction in 90-day mortality in patients with unplanned admission due to
sepsis and in those with planned admissions, both of which had a probability of benefit above 0.95. In
a continuous analysis, a “contamination” effect was present in most scenarios, with increasing
percentage of fluid being given as 0.9% saline before enrollment being associated with lower possible
benefits of receiving balanced solutions during ICU stay. In no scenario there was an above 0.90
probability that balanced solutions were associated with at least one day more alive and free of KRT
at 28 days.
The present analysis expands the results of the BaSICS trial by investigating potential sources
of heterogeneity in treatment effect using relevant effect modifiers that were not fully considered in
the trial main analyses. One key factor is baseline fluid use before enrollment in the trial; more than
60% of all patients in BaSICS used fluid at baseline which were not protocolized and varied across
sites. It is conceivable that both fluid use and admission type may interact and modulate the response
to different types of fluid in the ICU. Our results can, therefore, be interpreted as following: the most
promising signal for potential benefit appeared in patients that exclusively received balanced solutions
before enrollment; when further dividing this group according to admission type, the effect seemed to
be mostly associated with unplanned admission due to sepsis, being the results in other admission
types sensitive to priors or to sensitivity analyses. Results in the sepsis subgroup that exclusively
received balanced solutions before enrollment were insensitive to different priors; in fact, even when a
pessimistic prior that simulated a low (0.15) probability of benefit for the intervention was considered,
the posterior probability of benefit of balanced solutions was above 0.95. This is in contrast with the
signal for benefit in patients with planned admission, where results were more sensitive to different
priors (probability of benefit decayed from 0.99 for the optimistic prior to 0.85 for pessimistic prior),
suggesting that high uncertainty remains [11]. In all subgroups, increasing the percentage of fluid
infused as 0.9% saline before enrollment appeared to modulate the results, with higher percentages
blunting benefits of balanced solutions specially on planned admissions and in septic patients.
Our results appear to be consistent with a previous secondary analysis of patients with sepsis
in a prior trial that found that fluid therapy prior to ICU admission modified the effect of the fluid type
assigned by the trial on clinical outcomes [3,4], which motivated a suggestion for use of balanced
solutions in septic patients issued by the most recent Surviving Sepsis Guidelines [12]. The precise
explanation for our findings remain not completely elucidated. Chloride and hyperchloremia are
usually considered to be on the mediation pathway for the effect of balanced solutions. However,
0.9% saline use is not the only culprit for hyperchloremia in critical illness and in the postoperative
period [13,14]. In fact, the association between chloride infusion and changes in strong ion difference
in critically ill patients, although significant, may be of small magnitude (Figure E3 and [15]). . It is
unclear how use of 0.9% saline may blunt potential benefits of balanced solutions, although one can
hypothesize that abrupt increases on chloride causing fast reductions in strong ion difference and
(even if transitory) acidosis might be involved. A single serum chloride measurement may not reflect
all changes in electrolyte values that occur after fluid expansion because of redistribution. . The acid-
base profile in sepsis is usually characterized by an important decrease in albumin and increase in
chloride [16]. Due to severe inflammation and reduction in albumin synthesis [16], it could be
hypothesized that septic patients might be more sensitive to external chloride load, for example. This
association was not clear from baseline data available (Figure E2); however, frequency of
hyperchloremia (serum chloride >110 mEq/L) in the subgroup of patients that had chloride data
available was high and sustained (if not increased) during the first 3 days of the trial in the subgroup
of septic patients receiving 0.9% saline, while the frequency decreased in group receiving balanced
solutions (Figure E6). The group of planned admissions that received balanced solution before
enrollment and that was randomized to received balanced solutions also had a low frequency of
measured hyperchloremia. Although these findings might corroborate with the chloride hypothesis,
they are limited to only part of the included patients and were not incorporated in a formal mediation
analysis. Overall, one of the most important conclusions from our data is that confining a trial’s
intervention to a specific location (in our case, the intensive care unit), may increase the chance of
yielding neutral results when intervention is time sensitive and occurs irrespectively of patient’s locale
(as is the case of fluids, antibiotics, or many other interventions in acutely ill patients).
Bayesian investigations of heterogeneity in treatment effects are a promising tool and should
ideally be planned before trial initiation [17]. Investigating heterogeneity of treatment effect, specially
in the context of a neutral trial, is challenging [18,19]. If overall results are neutral and there is sign of
benefit in a specific subgroup, this suggests harm has occurred in other subgroup(s) of patients. None
of our subgroups had undeniable harm of balanced solutions. However, exclusion of traumatic brain
injury patients increased the overall trial probability of benefit, which may explain at least part of the
global neutral results in the presence of subgroups of patients that may have benefited from balanced
solutions. This was also noticeable on the continuous assessment of percentage use of fluid infused as
0.9% saline analysis, where exclusion of traumatic brain injury patients hampered the signal of harm
for balanced solutions seen at unplanned admissions without sepsis (Figure E18). Future endeavors,
including an ongoing individual patient metanalysis between SPLIT [209], SMART [1], BaSICS [2]
and PLUS trial [21] (PROSPERO CRD42022299282), will provide further information for different
This study has several limitations. First, it is a secondary post hoc analysis of a large
randomized controlled trial and should be therefore be seen as exploratory, although they are aligned
with other subgroup reports from the SMART trial. Second, the main analysis was performed by
arbitrarily classifying patients into 4 groups according to the use or not of each type of fluid before
enrollment while considering admission types groups that are largely heterogeneous; however, a
continuous analysis suggested that there was a continuous “contamination” effect of increasing the
percentage of fluid infused as saline before enrollment and a reduction in the potential benefits of
receiving balanced solutions in the ICU. Use of fluid before enrollment was a core variable in BaSICS
case report form; however, due to the pragmatic nature of the trial we did not monitor source
documents for all enrolled patients and it is conceivable that some misclassification may have
occurred. Around 60% of all patients had documented fluid use before enrollment, which is
compatible with data from the SMART trial [1]. Third, as with all Bayesian analyses, priors were used
in the analyses; although we have used priors previously suggested, prior selection can always be
considered subjective [6]. Forth, we made no distinction on the type of balanced solution before
enrollment. Finally, the Bayesian model created was heavily sampled. Bayesian models may be,
however, less sensitive to multiple comparison [22,23] and our simulations suggest that even in the
context of multiple comparisons the probability of obtaining a high probability of benefit (arbitrarily
set as above 0.90) was very low for the main group of interest. .
Conclusion
There is a high probability that balanced solution use in the ICU reduces 90-day mortality in
patients that exclusively received balanced fluids before trial enrollment, specially in the subgroup of
References
1. Semler MW, Self WH, Wanderer JP, Ehrenfeld JM, Wang L, Byrne DW, Stollings JL, Kumar AB,
Hughes CG, Hernandez A, Guillamondegui OD, May AK, Weavind L, Casey JD, Siew ED, Shaw
AD, Bernard GR, Rice TW; SMART Investigators and the Pragmatic Critical Care Research Group.
Balanced Crystalloids versus Saline in Critically Ill Adults. N Engl J Med. 2018 Mar 1;378(9):829-
839. doi: 10.1056/NEJMoa1711584. Epub 2018 Feb 27. PMID: 29485925; PMCID: PMC5846085.
2. Zampieri FG, Machado FR, Biondi RS, Freitas FGR, Veiga VC, Figueiredo RC, Lovato WJ,
Amêndola CP, Serpa-Neto A, Paranhos JLR, Guedes MAV, Lúcio EA, Oliveira-Júnior LC, Lisboa
TC, Lacerda FH, Maia IS, Grion CMC, Assunção MSC, Manoel ALO, Silva-Junior JM, Duarte P,
Soares RM, Miranda TA, de Lima LM, Gurgel RM, Paisani DM, Corrêa TD, Azevedo LCP, Kellum
JA, Damiani LP, Brandão da Silva N, Cavalcanti AB; BaSICS investigators and the BRICNet
members. Effect of Intravenous Fluid Treatment With a Balanced Solution vs 0.9% Saline Solution
on Mortality in Critically Ill Patients: The BaSICS Randomized Clinical Trial. JAMA. 2021 Aug
PMC8356144.
3. Jackson KE, Wang L, Casey JD, Bernard GR, Self WH, Rice TW, Semler MW; SMART
Investigators and the Pragmatic Critical Care Research Group. Effect of Early Balanced Crystalloids
4. Brown RM, Wang L, Coston TD, Krishnan NI, Casey JD, Wanderer JP, Ehrenfeld JM, Byrne DW,
Stollings JL, Siew ED, Bernard GR, Self WH, Rice TW, Semler MW. Balanced Crystalloids versus
Saline in Sepsis. A Secondary Analysis of the SMART Clinical Trial. Am J Respir Crit Care Med.
PMC6909845.
5. Zampieri FG, Azevedo LCP, Corrêa TD, Falavigna M, Machado FR, Assunção MSC, Lobo SMA,
Dourado LK, Berwanger O, Kellum JA, Brandão N, Cavalcanti AB; BaSICS Investigators and the
BRICNet. Study protocol for the Balanced Solution versus Saline in Intensive Care Study (BaSICS): a
factorial randomised trial. Crit Care Resusc. 2017 Jun;19(2):175-182. PMID: 28651514.
6. Zampieri FG, Casey JD, Shankar-Hari M, Harrell FE Jr, Harhay MO. Using Bayesian Methods to
Augment the Interpretation of Critical Care Trials. An Overview of Theory and Example Reanalysis
of the Alveolar Recruitment for Acute Respiratory Distress Syndrome Trial. Am J Respir Crit Care
PMC7924582.
7. R Core Team (2021). R: A language and environment for statistical computing. R Foundation for
2021
8. Bürkner PC. brms: An R Package for Bayesian Multilevel Models Using Stan. Journal of Statistical
9. Kay M. tidybayes: Tidy Data and Geoms for Bayesian Models. 10.5281/zenodo.1308151 , R
10. Self WH, Semler MW, Wanderer JP, Wang L, Byrne DW, Collins SP, Slovis CM, Lindsell CJ,
Ehrenfeld JM, Siew ED, Shaw AD, Bernard GR, Rice TW; SALT-ED Investigators. Balanced
Crystalloids versus Saline in Noncritically Ill Adults. N Engl J Med. 2018 Mar 1;378(9):819-828. doi:
11. Bampoe S, Odor PM, Dushianthan A, Bennett-Guerrero E, Cro S, Gan TJ, Grocott MP, James
MF, Mythen MG, O'Malley CM, Roche AM, Rowan K, Burdett E. Perioperative administration of
buffered versus non-buffered crystalloid intravenous fluid to improve outcomes following adult
surgical procedures. Cochrane Database Syst Rev. 2017 Sep 21;9(9):CD004089. doi:
12. Evans L, Rhodes A, Alhazzani W, Antonelli M, Coopersmith CM, French C, Machado FR,
Mcintyre L, Ostermann M, Prescott HC, Schorr C, Simpson S, Wiersinga WJ, Alshamsi F, Angus
DC, Arabi Y, Azevedo L, Beale R, Beilman G, Belley-Cote E, Burry L, Cecconi M, Centofanti J, Coz
Sprung CL, Welte T, Zimmerman J, Levy M. Surviving Sepsis Campaign: International Guidelines
for Management of Sepsis and Septic Shock 2021. Crit Care Med. 2021 Oct 4. doi:
13. McIlroy DR, Murphy D, Shotwell MS, Bhatia D. Peak Serum Chloride and Hyperchloremia in
Patients Undergoing Cardiac Surgery Is Not Explained by Chloride-Rich Intravenous Fluid Alone: A
Post-Hoc Analysis of the LICRA Trial. J Cardiothorac Vasc Anesth. 2021 May;35(5):1321-1331. doi:
14. Yessayan L, Neyra JA, Canepa-Escaro F, Vasquez-Rios G, Heung M, Yee J; Acute Kidney Injury
in Critical Illness Study Group. Effect of hyperchloremia on acute kidney injury in critically ill septic
patients: a retrospective cohort study. BMC Nephrol. 2017 Dec 2;18(1):346. doi: 10.1186/s12882-
15. Zampieri FG. Chapter 66 - Metabolic Acidosis. In Critical Care Nephrology, 3rd Edition, 2017.
16. Noritomi DT, Soriano FG, Kellum JA, Cappi SB, Biselli PJ, Libório AB, Park M. Metabolic
acidosis in patients with severe sepsis and septic shock: a longitudinal quantitative study. Crit Care
17.
18. Harrell, F. Viewpoints on Heterogeneity of Treatment Effect and Precision Medicine. Available
19. Rothwell PM, Mehta Z, Howard SC, Gutnikov SA, Warlow CP. Treating individuals 3: from
subgroups to individuals: general principles and the example of carotid endarterectomy. Lancet. 2005
Buffered Crystalloid Solution vs Saline on Acute Kidney Injury Among Patients in the Intensive Care
Unit: The SPLIT Randomized Clinical Trial. JAMA. 2015 Oct 27;314(16):1701-10. doi:
Investigators and the Australian and New Zealand Intensive Care Society Clinical Trials Group.
Balanced Multielectrolyte Solution versus Saline in Critically Ill Adults. N Engl J Med. 2022 Jan 18.
22. Berry DA, Hochbergb Y. Bayesian perspectives on multiple comparisons. Journal of Statistical
23. Sjölander, A., Vansteelandt, S. Frequentist versus Bayesian approaches to multiple testing. Eur J
Figure Legends:
Figure 1 – Forest plot of odds ratio of balanced solution versus 0.9% saline for mortality and their
respective 89% and 95% credible interval. 89% credible intervals are shown in bold lines and 95%
credible intervals are shown in light gray. All results shown are for the neutral prior. All results were
Figure 2 - Conditional distributions of the odds ratio for balanced solutions versus saline according to
fluid use group before enrollment (panels) and admission types (lines within panels). The probabilities
of benefit (OR<1) are highlighted in blue and the probability of harm in red.
Figure 3 – Results for the continuous assessment of fluid use before enrollment according to
percentage infused as 0.9% saline for (A) Unplanned admissions with sepsis, (B) Unplanned
admissions without sepsis, and (C) Planned admissions. Each facet panel represents a hypothetical
total volume of fluid received prior to enrollment (1000, 2000, 3000 and 4000 mL), and the x axis
represents the percentage of fluid volume infused as saline (from 0 to 1, at 0.1 steps). The boxplots
represent the model’s conditional predicted mortality probability at that specific combination of total
volume of fluid before enrollment and the percentage which was infused as 0.9% saline. The benefit
of balanced solution seems to be more pronounced when the percentage of fluid infused as 0.9%
saline before enrollment was low, specially up to 0.20 for unplanned admissions. For unplanned
admissions without sepsis, there was a trend towards higher mortality on patients randomized to
balanced solutions (B panels), which was importantly reduced when traumatic brain injury patients
were excluded (see Figure E6 in the online data supplement). For patients with planned admission,
differences were more pronounced only at higher volumes of fluid used before enrollment (C, bottom
left panel).
Tables
Table 1 – Patient characteristics according to fluid use before enrollment and randomization arm.
Table 2- Results for the primary endpoint in the Bayesian model following the neutral moderate
skeptical† - Primary group of interest; * Primary subgroup of interest. 89% credible intervals are
Difference in 90-
Fluid Group day mortality Odds Ratio for 90-day
P(OR P(OR<1/1.25) P(1/1.1<OR<1
Admission Type before (Balanced – Saline); mortality, median (89% CrI)
median (89% CrI) <1) [P(OR<0.8)] .1)
Enrollment [95% CrI]
[95% CrI]
Figure 1
Figure 2
Figure 3
(BaSICS) Study
Tables and Figures are shown in the order of appearance in the text
Table E1 – Prior definitions: We generally followed prior definitions for reanalysis based on Zampieri FG et al (Am J
Respir Crit Care Med. 2021 Mar 1;203(5):543-552. doi: 10.1164/rccm.202006-2381CP.). BaSICS was designed to
detect approximately something between 3.5-4% absolute risk reduction from an absolute 35% mortality (the very
initial draft considered 4% reduction for 35% baseline risk, which was subsequently changed to 3.5% reduction before
trial begin in one of the first protocol versions. We considered, for simplicity, that BaSICS aimed to detect something
close to a 4% reduction in absolute mortality considering a 35% baseline event rate, that is, an odds ratio of ~0.83 (-
0.182 on log scale). The optimistic prior was therefore set as a normal prior on log scale with mean set at the effect size
BaSICS was designed to detect (-0.182) and with a sufficiently large standard deviation so that 0.15 of probability of
harm remained. The pessimistic prior is the opposite of the optimistic prior (mean 0.182, same standard deviation), so it
acknowledges 0.15 probability of benefit. The neutral prior is set as defined on the original paper so that is is centered
as absence of effect (OR = 1; log[OR] = 0) and that 0.95 of all probability mass is contained between OR of 0.5 and 2.0.
Age, median [IQR] 64 (50, 75) 63 (52, 74) 62 (49, 73) 63 (52, 71)
Female, n(%) 1,521 (46%) 1,438 (45%) 909 (43%) 771 (41%)
Admission Type
Unplanned, no sepsis 1,789 (53%) 622 (19%) 802 (38%) 315 (17%)
Unplanned, sepsis 876 (27%) 332 (10%) 514 (25%) 178 (20%)
SOFA, median [IQR] 4.0 (2.0, 7.0) 4.0 (2.0, 6.0) 5.0 (3.0, 7.0) 4.0 (3.0, 6.0)
Creatinine, median [IQR] 1.06 (0.79, 1.55) 0.97 (0.75, 1.28) 1.00 (0.74, 1.50) 0.92 (0.70, 1.21)
90-day mortality 1260 (37.5) 573 (17.9) 686 (32.0) 299 (16.0)
Figure E1 – Volume of fluid use before enrollment according to admission type (panels), and fluid subgroup (x-axis).
Figure E2 – Association between volume of saline before enrollment infused and chloride serum levels at admission
(data available for 4,121 patients), stratified according to admission type. Note that the association between volume of
saline use and serum chloride at admission, albeit present, is of small magnitude and seems to be less pronounced in
Figure E3 – Mortality according to group of fluid use before enrollment and admission type.
Figure E4 – Mean and 95% confidence interval for chloride values at baseline and days 1, 2, and 3 for patients with
unplanned admission due to sepsis according to fluid use before enrollment. Values were available for [n(%)]: 596
(0.31), 698 (0.37), 632 (0.34), and 534 (0.28) patients, respectively.
Figure E5 – Mean and 95% confidence interval for chloride values at baseline and days 1, 2, and 3 for patients with
unplanned admission not due to sepsis according to fluid use before enrollment. Values were available for [n(%)]: 1,238
(0.35), 1,422 (0.40), 1,253 (0.35), and 1,103 (0.31) patients, respectively.
Figure E6 – Mean and 95% confidence interval for chloride values at baseline and days 1, 2, and 3 for patients with
planned admission according to fluid use before enrollment. Values were available for [n(%)]: 2,287 (0.45), 2,169
Figure E7 – Percentage of patients developing hyperchloremia (serum chloride >110 mEq/L) in the subgroup of 294
patients with unplanned admission due to sepsis with chloride levels available during the first 3 days according to
intervention (panel rows) and use of fluid before enrollment (columns). Note that ~40% of patients that received only
Balanced Solutions or 0.9% Saline before enrollment (first column) had hyperchloremia at enrollment; however,
percentage of patients hyperchloremia frequency reduced over time in the group of patients randomized to receive
Balanced Solutions in the ICU, while it remained constant in those randomized to 0.9% Saline.
Figure E8 – Percentage of patients developing hyperchloremia in the subgroup of 620 patients with unplanned
admission not due to sepsis with chloride levels available during the first 3 days according to intervention (panel rows)
Figure E9 – Percentage of patients developing hyperchloremia in the subgroup of 766 patients with planned admission
with chloride levels available during the first 3 days according to intervention (panel rows) and use of fluid before
enrollment (columns). Note how the percentage of patients with hyperchloremia is lower in patients randomized to
received balanced solutions and that received only balanced solutions prior to enrollment. Similarly, percentage of
patients with hyperchloremia decreased in patients randomized to received balanced solutions in the subgroup of
Table E3 - Results for the primary endpoint in the Bayesian model following the optimistic moderate strength prior.
Difference in 90-
Odds Ratio for
Fluid Group day mortality
90-day mortality,
Admission Type before (Balanced – Saline); P(OR<1) P(OR<1/1.25) P(1/1.1<OR<1.1)
median (89%
Enrollment median (89% CrI)
CrI) [95% CrI]
[95% CrI]
Table E4 - Results for the primary endpoint in the Bayesian model following the pessimistic moderate strength prior.
Difference in 90-
Odds Ratio for
Fluid Group day mortality
90-day mortality,
Admission Type before Balanced – Saline; P(OR<1) P(OR<1/1.25) P(1/1.1<OR<1.1)
median (89%
Enrollment median (89% CrI)
CrI) [95% CrI]
[95% CrI]
Figure E10 – Conditional probability of death according to fluid use before enrollment for the neutral moderate
strength prior according to fluid use before enrollment. (A) Only balanced solution used before enrollment; (B) Only
0.9% Saline used before enrollment; (C) Mixed fluid use; (D) No fluid use. Inside each panel the probability of death
according to randomization arm (fluid use after enrollment - balanced solution or 0.9% saline) are shown on the left,
and the difference between conditional probabilities in the plot on the right. The blue shaded are marks the probability
that mortality is lower in the balanced solution randomization arm. See Table 2 for numeric values. Values shown in
figure for the difference are median and 89% credible interval.
Figure E11 - Conditional probability of death according to admission type for the neutral moderate strength prior
according to admission type. (A) Unplanned admissions with sepsis; (B) Unplanned admissions without sepsis; (C)
Planned admissions. Inside each panel the probability of death according to randomization arm (fluid use after
enrollment - balanced solution or 0.9% saline) are shown on the left, and the difference between conditional
probabilities in the plot on the right. The blue shaded are marks the probability that mortality is lower in the balanced
solution randomization arm. See Table 2 for numeric values. Values shown in figure for the difference are median and
Figure E12 – Conditional probability of death according to admission type for patients that received a balanced
solutions before enrollment, considering the the neutral moderate strength prior. (A) Unplanned admissions with sepsis;
(B) Unplanned admissions without sepsis; (C) Planned admissions. Values shown in figure for the difference are
Figure E13 – Conditional probability of death according to admission type for patients that received a mix of balanced
solutions and 0.9% saline before enrollment, considering the the neutral moderate strength prior. (A) Unplanned
admissions with sepsis; (B) Unplanned admissions without sepsis; (C) Planned admissions. Values shown in figure for
Figure E14 – Conditional probability of death according to admission type for patients that received no fluids before
enrollment, considering the the neutral moderate strength prior. (A) Unplanned admissions with sepsis; (B) Unplanned
admissions without sepsis; (C) Planned admissions. Values shown in figure for the difference are median and 89%
credible interval.
Figure E15 – Conditional probability of death according to admission type for patients that received only 0.9% saline
before enrollment, considering the the neutral moderate strength prior. (A) Unplanned admissions with sepsis; (B)
Unplanned admissions without sepsis; (C) Planned admissions. Values shown in figure for the difference are median
Table E5 - Results for the primary endpoint in the Bayesian model following flat priors.
Difference in 90-
Odds Ratio for
Fluid Group day mortality
90-day mortality,
Admission Type before (Balanced – Saline); P(OR<1) P(OR<1/1.25) P(1/1.1<OR<1.1)
median (89%
Enrollment median (89% CrI)
CrI) [95% CrI]
[95% CrI]
Table E6 - Results for the primary endpoint in the Bayesian model following the neutral moderate skeptical prior
Unplanned, sepsis Mix 0.08 (-0.03,0.20) 1.43 (0.87,2.37) 0.12 0.03 0.12
Unplanned, no sepsis Mix 0.02 (-0.06,0.11) 1.11 (0.72,1.71) 0.35 0.11 0.27
Secondary Analyses
before enrollment is lost when only four groups are considered. The first sensitivity analysis aimed at evaluating a
potential “dose-contamination” effect increasing the percentage of fluid infused as 0.9% saline before enrollment. This
analysis was performed using a generalized mixed additive model with a tensor effect between volume of saline before
enrollment, volume of balanced solution before enrollment, and intervention (balanced solution versus 0.9% saline).
This analysis therefore encompasses both the volume of each fluid used and their interaction at the treatment levels. We
applied neutral normal priors with mean 0 and standard deviation of 1 to regularize the posteriors and built three
models, one for each admission type (unplanned with sepsis, unplanned without sepsis, and planned. The syntax for this
model was:
The three resulting models were sub-sequentially sampled for possible combinations of total volume of fluid
infused and percentage used as saline, with 4,000 samples of each possible combinations sampled and shown in the
form of boxplots. For clarity, results are displayed as (1) conditional probabilities of death according to randomization
arm for combinations of total fluid use before enrollment and percentage of fluid before enrollment infused as 0.9%
saline (shown in the manuscript, Figure 4), and (2) conditional death probabilities according to study arm for fixed
volumes of 0.9% saline and balanced solution use before enrollment. Results for (1) are based on potential scenarios
where total fluid use before enrollment was of 1,000, 2,000, 3,000 or 4,000 mL with varying percentages of fluids
infused as saline. Results of the second presentation approach are shown in Figures E9, E10, and E11, below. A
sensitivity analysis on unplanned admissions without sepsis was made excluding traumatic brain injury patients and is
Figure E16 – Predicted mortality according to combinations of volumes of saline infused before enrollment (panels A
to D) and volumes of balanced solution infused before enrollment (subpanels) according to randomization arm (colors)
for unplanned admissions with sepsis. Note how differences between both groups become smaller after 1L of saline has
Figure E17 – Predicted mortality according to combinations of volumes of saline infused before enrollment (panels A
to D) and volumes of balanced solution infused before enrollment (subpanels) according to randomization arm (colors)
Figure E18 – Predicted mortality according to combinations of volumes of saline infused before enrollment (panels A
to D) and volumes of balanced solution infused before enrollment (subpanels) according to randomization arm (colors)
Figure E19 - Results for the continuous assessment of fluid use before enrollment according to percentage infused as
0.9% saline for unplanned admissions without sepsis excluding traumatic brain injury patients. Note how the
association between balanced solution randomization arm and higher mortality shown in Figure 4B is greatly reduced.
Predicted probabilities from the frequentist model are shown in the Figure E below.
P values for the main model can be cumbersome to present and interpret due to issues in providing p values in
situations where the number of degrees of freedom is unclear (see reference 2), which is the case for the present
frequentist model. The raw output for the frequentist model with P values extracted from Wald’s test are provided as
requested during peer review, but should be cautiously interpreted. Other methods of p-value estimation including those
based on Satterthwaite or Kenward-Roger methods for denominator degrees of freedom are yet not implemented on the
{lmerTest} R package for glmer models.
Table E8 - Coefficients of the frequentist model. P-values were obtained from Wald’s test, default in glmer() function
on {lme4} package. volgroup refers to the group of fluid use before enrollment (mix = Mixed, nf = No fluid, saline =
Saline Only, the level ‘balanced’ which represents Balanced Only was used for reference); admtype refers to admission
type (no_sepsis = Unplanned, not sepsis, sepsis = Unplanned, sepsis, the level ‘planned’ which represents planned
admission was used as reference); intervention refers to the randomization arm (intervention = Intervention; control
level was used as reference).
Estimate Std. Error z value Pr(>|z|)
(Intercept) -1.78 0.13 -13.32 0.00
volgroupmix -0.15 0.17 -0.9 0.37
volgroupnf 0.35 0.18 1.97 0.05
volgroupsaline 0.11 0.18 0.61 0.54
admtypeno_sepsis 1.13 0.16 6.9 0.00
admtypesepsis 1.97 0.19 10.21 0.00
interventionintervention -0.21 0.14 -1.5 0.13
volgroupmix:admtypeno_sepsis -0.37 0.29 -1.29 0.20
volgroupnf:admtypeno_sepsis -0.3 0.23 -1.32 0.19
volgroupsaline:admtypeno_sepsis -0.26 0.24 -1.07 0.29
volgroupmix:admtypesepsis -0.46 0.33 -1.4 0.16
volgroupnf:admtypesepsis -0.57 0.26 -2.2 0.03
volgroupsaline:admtypesepsis -0.46 0.27 -1.69 0.09
volgroupmix:interventionintervention 0.16 0.23 0.69 0.49
volgroupnf:interventionintervention 0.11 0.25 0.44 0.66
volgroupsaline:interventionintervention 0.11 0.24 0.45 0.65
admtypeno_sepsis:interventionintervention 0.01 0.23 0.05 0.96
admtypesepsis:interventionintervention -0.28 0.28 -1.00 0.32
volgroupmix:admtypeno_sepsis:interventionintervention 0.27 0.4 0.68 0.50
volgroupnf:admtypeno_sepsis:interventionintervention 0.07 0.33 0.23 0.82
volgroupsaline:admtypeno_sepsis:interventionintervention 0.21 0.34 0.61 0.54
volgroupmix:admtypesepsis:interventionintervention 0.71 0.46 1.53 0.13
volgroupnf:admtypesepsis:interventionintervention 0.26 0.37 0.7 0.48
volgroupsaline:admtypesepsis:interventionintervention 0.5 0.39 1.3 0.19
References:
1. Russell V. Lenth (2021). emmeans: Estimated Marginal Means, aka Least-Squares Means. R package version 1.7.1-
1. https://CRAN.R-project.org/package=emmeans
2. Douglas Bates, Martin Maechler, Ben Bolker, Steve Walker (2015). Fitting Linear Mixed-Effects Models Using
lme4. Journal of Statistical Software, 67(1), 1-48. doi:10.18637/jss.v067.i01.
with 0 being attributed to patients that either died or required at least one session of KRT in the period. Days alive and
free of kidney replacement therapy up to 28 days was assessed using a Bayesian beta binomial model with the same
adjustment as for the main model. We used non-informative priors (normal prior with mean 0 and standard deviation of
1 for all coefficients except for standard deviation for the intercept which was set as a Student t distribution with mean
0, standard deviation of 2.5 at 3 degrees of freedom). Results are presented with median and 89% credible interval for
the difference between days alive and free of renal replacement therapy in groups obtained by posterior sampling as as
probability that balanced solutions increased number of days alive and free of kidney replacement therapy at 28 days by
Unplanned, no Balanced
0.80 [-0.66,2.24] 0.41
sepsis Only
Unplanned, no Mix
-2.12 [-4.10,-0.11] 0.01
sepsis
Unplanned, no No Fluids
-0.11 [-1.01,0.78] 0.02
sepsis
Planned Balanced
0.38 [-0.14,0.90] 0.03
Only
hypothesis.
Readers may feel a little concerned about the number of queries in the posterior; specifically, readers may have
the feeling that “with so many analysis something would have to turn out positive”. Our results are exploratory, and we
have only one main model for the primary outcome and all analyses are based on that model. Spurious findings are
always possible. Even when we say that there is, say, 0.94 probability of benefit on a given scenario, there is still 6%
probability we got the signal wrong (and the intervention is harmful). Interpretation of Bayesian methods should be
made on the context of the posterior distribution of the effect size given the priors and our data. Simulation studies can
In the code below, we run 2,500 simulations on 2,500 hypothetical results of BaSICS trial. The random site
effect was remove for simplicity (and to spare some computational time; each simulation takes around 10 minutes to
run). We first create a data frame with the same number of patients in each group (admission type and fluid use before
enrollment). Intervention (Plasma Lyte of 0.9% saline) is sampled from a binomial distribution of probability 0.5 in
each subgroup and mortality within these 12 subgroups is thereafter sample from a binomial distribution with a
probability equal of the one observed in the main trial. This is made so that the randomization arm and death probability
are independent within each subgroup, but each subgroup still has a mortality similar to the main trial. In this scenario,
we expect that the median odds ratio for all subgroups are centered at 1 (absence of effect) and that for all scenarios the
probability of benefit is centered at 0.5, with lower probability mass on tails. We provide the code below for the reader.
####Load Packages####
library(tidyverse)
library(brms)
library(tidybayes)
#Function to create a simulated data frame that respects the same number of patients in each group as BaSICS
createdata<-function(x){
planejada<-data.frame(admtype=rep("Planejada",2248+1369+695+780),
volgroup=c(rep("Balanced Only",2248),
rep("Mix",1369),
rep("No fluids",695),
rep("Saline Only",780)),
death=c(rbinom(2248,1,0.101),
rbinom(1369,1,0.102),
rbinom(695,1,0.170),
rbinom(780,1,0.171))
volgroup=c(rep("Balanced Only",622),
rep("Mix",315),
rep("No fluids",1789),
rep("Saline Only",802)),
death=c(rbinom(622,1,0.312),
rbinom(315,1,0.270),
rbinom(1789,1,0.391),
rbinom(802,1,0.375))
volgroup=c(rep("Balanced Only",332),
rep("Mix",178),
rep("No fluids",876),
rep("Saline Only",514)),
death=c(rbinom(332,1,0.455),
rbinom(178,1,0.421),
rbinom(876,1,0.506),
rbinom(514,1,0.490))
teste<-bind_rows(planejada,naoplanejada,naoplanejadasepse)
return(teste)
pimpme_all<-function(mbrms=m1n){
step1<-tdata%>%add_epred_draws(mbrms)
step2<-step1%>%
ungroup()%>%
select(-.row,-.chain,-.iteration,-.draw,-volgroup,-admtype)%>%
group_split(intervention)
step3<-data.frame(pdiff=((step2[[1]]$.epred)-(step2[[2]]$.epred)),
r1<-step3%>%
dplyr::summarise(median_or=median(podds),
lower_odds=quantile(podds,probs=0.055),
upper_odds=quantile(podds,probs=0.945),
pbenefit=sum(podds<1)/n())
r1$admtype<-"All"
r1$volgroup<-"All"
return(r1[,c(5,6,1,2,3,4)])
pimpme<-function(mbrms=m1n){
step1<-tdata%>%add_epred_draws(mbrms)
step2<-step1%>%
ungroup()%>%
select(-.row,-.chain,-.iteration,-.draw)%>%
group_split(admtype,volgroup,intervention)
step3<-list()
for (i in seq(2,24,by=2)) {
k=i/2
step3[[k]]=data.frame(admtype=step2[[i]]$admtype,
volgroup=step2[[i]]$volgroup,
pdiff=((step2[[(i-1)]]$.epred)-(step2[[(i)]]$.epred)),
(step2[[(i)]]$.epred))) ))
step3<-bind_rows(step3)
r2<-step3%>%group_by(admtype,volgroup)%>%
dplyr::summarise(median_or=median(podds),
lower_odds=quantile(podds,probs=0.055),
upper_odds=quantile(podds,probs=0.945),
pbenefit=sum(podds<1)/n())
return(r2)
tdata<-expand_grid(admtype=levels(as.factor(createdata()$admtype)),
intervention=levels(as.factor(createdata()$intervention)),
volgroup=levels(as.factor(createdata()$volgroup)))
mbas<-brm(death~volgroup*admtype*intervention,
prior=nprior)
#Simulation function
simme<-function(x){
x1<-update(mbas,newdata = createdata(),prior=nprior,chains=4,cores=4)
rm(x1)
return(volta)
rlist<-list()
for(i in 1:2500){
print(i)
rlist[[i]]<-simme()
#Simulation should return a list with 2,500 entries with all the results. These can be queried as desired.
As expected, the distribution of odds ratio (A) for the whole sample and the probability of benefit for the whole
sample (B) are centered at 1 and 0.5, respectively. This is a proof that simulation succeeded on producing neutral
results:
Distribution of odds ratios for all 12 subgroups is shown below, and is also centered at 1. Note that since each
subgroup has a much smaller sample than the original trial the distributions are slightly more flatted. Values of OR <0.8
Distribution of probabilities of benefit are also shown, with values for probability of benefit above 0.90
highlighted in blue:
From a practical perspective, one could query how frequently the simulations yield given probabilities as we
1. We report that overall probability of benefit of balanced solutions for patients that received only balanced
solutions across all admission types was 0.92. We saw similar results in ~0.065 of all simulations.
2. We report that overall probability of benefit of balanced solutions for patients with unplanned admissions due
to sepsis that received only balanced solutions before enrollment was 0.96; We saw similar results in ~0.003
of all simulations.
3. We report that overall probability of benefit of balanced solutions for patients with planned admission that
received only balanced solutions before enrollment was 0.97; We saw similar results in ~0.028 of all
simulations.
Therefore, the results above suggest that even in the context of a heavily queried sample Bayesian model, the
probability of finding a results as extreme as ours (for the main groups of interest) given that no effect was present was
The full code of this manuscript is available upon request at fgzampieri@gmail.com. We provide snippets
below that should be fully function under any data. There is one, single, main model for the primary endpoint:
90-day mortality ~ admission type * Fluid use before enrollment * Randomization arm + (1 | Enrolling ICU)
All that comes after is a consequence of this single model for the primary outcome. The philosophy of all we
2. Create a toy data set of all possible combinations between admission type, fluid use before enrollment and
intervention.
3. Query the model for 4,000 predicted probability values from the posteriors of each possible combinations
4. Use the predicted probabilities to generate relative risks, absolute risk reductions and odds ratios.
the concepts we used. We used the add_epred_draws function in {tidybayes} package to extract 4,000 samples of
possible scenarios from the posteriors. This specific R function extracts the probabilities from the Bayesian models
This is very similar to the approach of defining heterogeneity in treatment effects discussed in this manuscript:
enough. It even discusses the “do” operator defined by Judea Pearl in the context of estimating heterogeneity of
treatment effects. The reader should understand what we did as an attempt to measure counterfactual probabilities; that
is, probabilities that arouse from the same model while changing one of the predictors. The odds ratio obtained are,
consequently, odds ratios that are conditional to the models, priors and data. Therefore, all the effect sizes obtained are
the result of exploring and combining the predicted conditional probabilities for hypothetical combinations. For
example, for the effect size of balanced solutions for unplanned admissions due to sepsis and patients that only received
1. Sample the posteriors from the main model fixing admission type as septic patients and those that received
only balanced solutions and that were randomized to 0.9% saline. Let’s call this R1.
2. Sample the posteriors from the main model fixing admission type as septic patients and those that received
only balanced solutions and that were randomized to 0.9% saline. Let’s call this R2.
3. The relative risk is given by dividing the 4,000 samples in R2 by the 4,000 samples on R1; relative risk
reduction is given by R2 – R1 and odds ratio is given by [R2/(1-R2)]/[R1/(1-R1)].
4. We obtain a distribution of possible odds ratios (and relative risks, and absolute risk reductions, etc). We can
then obtain probability of benefit by counting the number of results that yield an odds ratio below 1 on that
sampling. The same process can be made for estimating probability of odds ratio being below 0.8 or within any
given range. Therefore, our results of “probability of benefit” are the probability of obtaining an odds ratio
#Loads packages
library(tidybayes)
library(brms)
We then proceed to defining priors and running the models for neutral, optimistic and pessimistic priors:
#Neutral prior
#Optimistic prior
#Pessimistic prior
m1n<-brm(death~volgroup*admtype*intervention+(1|icu),
prior=nprior)
m1o<-brm(death~volgroup*admtype*intervention+(1|icu),
prior=oprior)
m1p<-brm(death~volgroup*admtype*intervention+(1|icu),
prior=pprior)
Note that we used a very weak regularizing prior [N(0,1)] for all other categorical variables with the exception
After some time (which may take several minutes), models should run and then it is time to query the
posteriors. We now define some functions to perform all the queries we want for the whole population, specific fluid
uses before enrollment (regardless of admission type), specific admission types (regardless of fluid use before
enrollment), and combination subgroups of admission types and fluid use before enrollment. The codes are:
#Main result
result_all<-function(mbrms=m1n){
step1<-tdata%>%add_epred_draws(mbrms)
step2<-step1%>%
ungroup()%>%
select(-icu,-.row,-.chain,-.iteration,-.draw,-volgroup,-admtype)%>%
group_split(intervention)
step3<-data.frame(pdiff=((step2[[1]]$.epred)-(step2[[2]]$.epred)),
r1<-step3%>%
dplyr::summarise(median_diff=median(pdiff),
lower_diff=quantile(pdiff,probs=0.055),
upper_diff=quantile(pdiff,probs=0.945),
median_or=median(podds),
lower_odds=quantile(podds,probs=0.055),
upper_odds=quantile(podds,probs=0.945),
pbenefit=sum(podds<1)/n(),
pibenefit=sum(podds<0.8)/n(),
prope=sum(between(podds,1/1.1,1.1))/n())
return(list(r1,bind_rows(step2)))
#It is required to replace “grupo” for the group of fluid use before enrollment, namely
results_volgroup<-function(mbrms=m1n,grupo="Balanced Only"){
step1<-tdata%>%add_epred_draws(mbrms)
step2<-step1%>%
filter(volgroup==grupo)%>%
ungroup()%>%
select(-icu,-.row,-.chain,-.iteration,-.draw,-volgroup,-admtype)%>%
group_split(intervention)
step3<-data.frame(pdiff=((step2[[1]]$.epred)-(step2[[2]]$.epred)),
r1<-step3%>%
dplyr::summarise(median_diff=median(pdiff),
lower_diff=quantile(pdiff,probs=0.055),
upper_diff=quantile(pdiff,probs=0.945),
median_or=median(podds),
lower_odds=quantile(podds,probs=0.055),
upper_odds=quantile(podds,probs=0.945),
pbenefit=sum(podds<1)/n(),
pibenefit=sum(podds<0.8)/n(),
prope=sum(between(podds,1/1.1,1.1))/n())
return(list(r1,bind_rows(step2)))
#Planned = Planejada
step1<-tdata%>%add_epred_draws(mbrms)
step2<-step1%>%
filter(admtype==grupo)%>%
ungroup()%>%
select(-icu,-.row,-.chain,-.iteration,-.draw,-volgroup,-volgroup)%>%
group_split(intervention)
step3<-data.frame(pdiff=((step2[[1]]$.epred)-(step2[[2]]$.epred)),
r1<-step3%>%
dplyr::summarise(median_diff=median(pdiff),
lower_diff=quantile(pdiff,probs=0.055),
upper_diff=quantile(pdiff,probs=0.945),
median_or=median(podds),
lower_odds=quantile(podds,probs=0.055),
upper_odds=quantile(podds,probs=0.945),
pbenefit=sum(podds<1)/n(),
pibenefit=sum(podds<0.8)/n(),
prope=sum(between(podds,1/1.1,1.1))/n())
return(list(r1,bind_rows(step2)))
#Finally, there is one function to rule them all that extracts all possible combinations
results_subgroups<-function(mbrms=m1n){
step1<-tdata%>%add_epred_draws(mbrms)
step2<-step1%>%
ungroup()%>%
select(-icu,-.row,-.chain,-.iteration,-.draw)%>%
group_split(admtype,volgroup,intervention)
step3<-list()
for (i in seq(2,24,by=2)) {
k=i/2
step3[[k]]=data.frame(admtype=step2[[i]]$admtype,
volgroup=step2[[i]]$volgroup,
pdiff=((step2[[(i-1)]]$.epred)-(step2[[(i)]]$.epred)),
(step2[[(i)]]$.epred))) ))
step3<-bind_rows(step3)
#Sumários de benefícios
#Por comparação
r2<-step3%>%group_by(admtype,volgroup)%>%
dplyr::summarise(median_diff=median(pdiff),
lower_diff=quantile(pdiff,probs=0.055),
upper_diff=quantile(pdiff,probs=0.945),
median_or=median(podds),
lower_odds=quantile(podds,probs=0.055),
upper_odds=quantile(podds,probs=0.945),
pbenefit=sum(podds<1)/n(),
pibenefit=sum(podds<0.8)/n(),
prope=sum(between(podds,1/1.1,1.1))/n())
return(list(r2,step1))
In brief, all functions work by making predictions and comparisons with their counterfactual according to
randomization arm. There was possibly a way to merge all 3 functions on one, but we felt more comfortable debugging
with separate functions. Note that all the original code was made considering 89% credible intervals. The user can
change it if desired. We added 95% credible intervals for the main analysis during the peer review phase.
All functions returned both a data frame with predictions and a summarised version which was used to
populate the tables. We did that three times, one for each prior. We also did one forth time for the neutral prior, by
request.
Overall, that is the spirit of the main analysis. All figures, tables, etc, originated from these results. The reader
should keep in mind that this method, albeit technically correct in our opinion, is not the only way to summarise
contrasts from models in R. Both packages {emmeans} and {marginaleffects} can provide similar outputs (with nice
graphical summaries) and probably with less coding. R is an evolving language, and this code is going to be obsolete
soon. This code works with R 4.1.2, brms 2.16.3 and tidybayes v 3.0.1. These packages sometime changes their default
over time, so it is unpredictable if this will work for long. The reader can use the createdata() function on the simulation
code to play around with these models and results, if desired. If you read this far and believe this is useful, please let the