Restricted Versus Liberal Versus Goal-Directed Fluid Therapy for Non-vascular Abdominal Surgery: A Network Meta-Analysis and Systematic Review

Optimal perioperative fluid management is crucial, with over- or under-replacement associated with complications. There are many strategies for fluid therapy, including liberal fluid therapy (LFT), restrictive fluid therapy (RFT) and goal-directed fluid therapy (GDT), without a clear consensus as to which is better. We aimed to find out which is the more effective fluid therapy option in adult surgical patients undergoing non-vascular abdominal surgery in the perioperative period. This study is a systematic review and network meta-analysis (NMA) with node-splitting analysis of inconsistency, sensitivity analysis and meta-regression. We conducted a literature search of Pubmed, Cochrane Library, EMBASE, Google Scholar and Web of Science. Only studies comparing restrictive, liberal and goal-directed fluid therapy during the perioperative phase in major non-cardiac surgery in adult patients will be included. Trials on paediatric patients, obstetric patients and cardiac surgery were excluded. Trials that focused on goal-directed therapy monitoring with pulmonary artery catheters and venous oxygen saturation (SvO2), as well as those examining purely biochemical and laboratory end points, were excluded. A total of 102 randomised controlled trials (RCTs) and 78 studies (12,100 patients) were included. NMA concluded that goal-directed fluid therapy utilising FloTrac was the most effective intervention in reducing the length of stay (LOS) (surface under cumulative ranking curve (SUCRA) = 91%, odds ratio (OR) = -2.4, 95% credible intervals (CrI) = -3.9 to -0.85) and wound complications (SUCRA = 86%, OR = 0.41, 95% CrI = 0.24 to 0.69). Goal-directed fluid therapy utilising pulse pressure variation was the most effective in reducing the complication rate (SUCRA = 80%, OR = 0.25, 95% CrI = 0.047 to 1.2), renal complications (SUCRA = 93%, OR = 0.23, 95% CrI = 0.045 to 1.0), respiratory complications (SUCRA = 74%, OR = 0.42, 95% CrI = 0.053 to 3.6) and cardiac complications (SUCRA = 97%, OR = 0.067, 95% CrI = 0.0058 to 0.57). Liberal fluid therapy was the most effective in reducing the mortality rate (SUCRA = 81%, OR = 0.40, 95% CrI = 0.12 to 1.5). Goal-directed therapy utilising oesophageal Doppler was the most effective in reducing anastomotic leak (SUCRA = 79%, OR = 0.45, 95% CrI = 0.12 to 1.5). There was no publication bias, but moderate to substantial heterogeneity was found in all networks. In preventing different complications, except mortality, goal-directed fluid therapy was consistently more highly ranked and effective than standard (SFT), liberal or restricted fluid therapy. The evidence grade was low quality to very low quality for all the results, except those for wound complications and anastomotic leak.


Background
Perioperative fluid management is an important factor in perioperative care that contributes to long-term mortality and morbidity [1] and is frequently debated amongst perioperative physicians [2][3][4][5][6]. The three current options are restrictive fluid therapy (RFT), goal-directed fluid therapy (GDT) and liberal fluid 1 1 1 1 1 therapy (LFT). LFT assists with compensatory intravascular volume expansion, meets physiological requirements and compensates for blood loss, perioperative fasting and redistribution of third space. However, overhydration may lead to tissue oedema with poor wound healing, respiratory and cardiovascular complications and delayed recovery [1]. RFT aims for zero balance, less fluid volume given, with purported reduced perioperative complications [1] and shorter hospital stays. However, it may increase the risk of renal impairment and oliguria. GDT targets various measured hemodynamic variables to optimise hemodynamic status and oxygen delivery. Multiple randomised controlled trials (RCTs) and meta-analyses have been performed comparing the two of these three fluid therapies. Previous meta-analyses concluded that a restrictive approach may be beneficial [2][3][4][5]. However, since then, there have been new RCTs on fluid therapies, such as the Restrictive versus Liberal Fluid Therapy for Major Abdominal Surgery (RELIEF) trial in 2018, which concluded that a liberal approach did not increase complication rates and a restrictive approach resulted in increased risks of renal impairment [7]. There is only one updated meta-analysis included up to the RELIEF trial [8], and other later RCTs have still not been analysed in a meta-analysis [8,9].
With the advancement in intraoperative monitoring, novel means of achieving GDT have become available. Zhao et al. [10] recently performed a network meta-analysis (NMA) that compared these different GDTs in patients undergoing all types of surgery but did not compare GDT with other fluid therapies. NMA is an emerging technique for comparing multiple interventions in a single analysis through pooling direct and indirect evidence [11]. It allows different interventions to be compared indirectly if they have also been compared to one common intervention in different RCTs. This study is the first to utilise NMA to compare the many subgroups of GDT with RFT, LFT and standard fluid therapies (SFTs) at the same time.

Objectives
We aimed to investigate which is the more effective fluid therapy strategy in adult surgical patients undergoing non-vascular abdominal surgery during the perioperative period. Is there a difference in the perioperative mortality, length of stay (LOS), complication rates and organ-specific complication rates? This study aims to compare the effectiveness of RFT versus LFT versus GDT in adult patients undergoing nonvascular abdominal surgery during the perioperative period using an NMA and systematic review of the available data.

Protocol and Registration
We utilised the Preferred Reporting Items for Systematic Reviews and Meta-Analysis Extension Statement for Reporting of Systematic Reviews Incorporating Network Meta-Analyses of Health Care Interventions (PRISMA-NMA) guideline [12]. This work was not registered on any systematic reviews registry.

Inclusion and Exclusion Criteria
Only RCTs comparing the effects of RFT, LFT and GDT during the perioperative phase in major non-cardiac surgery in adult patients were included. Studies of intraoperative and postoperative fluid therapy were included. To support the transitivity assumption in NMA, similar studies were selected [13]. Studies on fluid therapy in cardiothoracic surgery, orthopaedic surgery, neurosurgery and vascular surgery were excluded from the NMA but included in the systematic review. Trials on paediatric patients and obstetric patients were excluded. Only journal articles published in English were included. Articles that did not report the primary or secondary end points were excluded. Trials that focused on GDT with pulmonary artery catheters and venous oxygen saturation (SvO2), as well as those examining purely biochemical and laboratory end points, were excluded. Grey literature such as non-traditional articles including reports, audits, editorials, commentaries, conference reports and abstracts were excluded.

Information Sources and Literature Search
We conducted a literature search of Pubmed, Cochrane Library, EMBASE, Google Scholar and Web of Science. All searches were performed from December 2020 to January 2021. The databases were searched separately to allow mapping to relevant subject headings. The search strategy followed validated methods of the QUOROM statement and Cochrane collaboration [14,15]. We expanded the subject headings search to include all relevant subheadings but restricted to human studies and the English language. Search terms included multiple combinations of Medical Subject Headings (MeSH), such as fluid therapy, surgery, perioperative, operative, complications, restrictive, liberal, goal-directed, FloTrac, esophageal/oesophageal Doppler, pulse pressure variation, optimisation/optimization and others. We did not apply restrictions on publication status or date.

Study Selection and Data Collection Process
Two researchers (YX and AT) independently screened articles by their titles and abstracts to identify eligible studies. Eligible entries then proceeded to full-text review. The references of the identified articles were hand-searched to avoid missing relevant studies. Any new studies found this way had their reference list manually searched. YX and AT used a predesigned data abstraction form to record the trial characteristics, outcomes and information related to the quality of the trial. We scored each trial according to the Jadad scale [16]. Two adjudicators (LL and LW) helped resolve disagreement by consensus. The extracted data was then laid out in a systematic review table (Appendices).

Data Items
The following data were extracted from eligible studies: author name, publication year, study design (whether the randomised controlled trial was single centre or multicentre), surgery type, intervention type and control, intervention phase (preoperative, intraoperative, or postoperative), fluid protocols and hemodynamic monitors and patient number (intervention versus control). The primary outcomes were hospital mortality (both 30-day and 90-day mortality, if reported), length of stay (LOS) (days) and complication rate. When articles reported LOS in terms of median and interquartile range or range, or mean and 95% confidence intervals, the mean and standard deviation were calculated using validated formulas proposed by McGrath et al. [17], Wan et al. [18] and Higgins et al. [19]. The complication rate was reported as the number of patients having at least one complication to avoid double-counting patients and to ensure the event rate was less than the sample size. The secondary outcomes were complication severity (major or minor) and organ-specific complications, which included respiratory complications (respiratory failure, pulmonary oedema, pneumonia and pleural effusion), cardiac complications (cardiac failure, myocardial infarction and arrhythmia), wound infections (both deep and superficial were pooled into a single composite end point), renal complications (including any renal impairment and need for dialysis) and anastomotic leak.

Geometry of the Network
To avoid imbalanced distribution of studies per node and allow for meaningful comparisons between treatment options, we classified the studies based on the authors' intentions and labels. Following the current literature, we defined RFT, which aimed for zero to negative fluid balance, with minimal physiological maintenance and minimal preloading of intravenous fluid before induction [20]. LFT was defined according to Miller's Anesthesia [21]. There was no standard definition for SFT amongst different centres, and some studies did not specify it in their protocol. GDT was defined as a protocolised approach using hemodynamic monitors to direct volume replacement to achieve certain hemodynamic goals. We pooled interventions with only one or two studies into a single node (GDTOthers). As the pulse variability index and pulse pressure variation are based on similar concepts, these were pooled to reduce inconsistency (GDTPpv). The majority of GDT studies utilised oesophageal Doppler (GDTOD), Edwards Lifesciences Vigileo/FloTrac (GDTFlo) and lithium dilution cardiac output (LiDCO) (GDTLid). Studies with three arms were included. However, if two arms with the same GDT monitoring modality use different fluids, the data of both arms were pooled.

Risk of Bias Within Individual Studies
We used a modified Cochrane Collaboration Risk of Bias Tool (RoB2) to assess the risk of bias in randomised trials [19]. We assessed selection bias including random sequence generation and allocation concealment, performance bias such as blinding of participants and personnel, detection bias such as blinding of outcome assessment, attrition bias due to incomplete outcome data, and reporting bias due to selective reporting. The overall risk of bias was then assigned to each study and categorised as high risk, low risk or unclear.

Summary Measures
We estimated the comparative efficacy of each fluid therapy using mean difference (MD), odds ratio (OR) and 95% credible intervals (CrI). We then used the surface under cumulative ranking curve (SUCRA) to visualise the treatment hierarchy. For comparisons with significant inconsistency, we report the direct, indirect and network evidence with comments on the reliability of the evidence.

Methods of Statistical Analysis
We used RevMan 5.4 to produce a risk of bias assessment and summary as per the Cochrane guide [19]. For the NMA, we used R version 4.0.3 (R Foundation for Statistical Computing, Vienna, Austria), Rstudio version 1.4.1103, and the gemtc, rjags, dmetar and netmeta packages to perform the statistical analysis [22]. All statistical analysis was performed by researcher YX and corroborated by AT. We selected a Bayesian approach for better coverage and estimation of effects in terms of probabilities [23][24][25]. We performed Markov chain Monte Carlo (MCMC) simulations using a random effects model, and final estimates were based on stable posterior sampling after initial iterations were discarded. We assessed the simulations for convergence using trace plots and the Brooks-Gelman-Rubin statistic and assessed the density plots for normal distribution. The potential scale reduction factor (PSRF) was below 1.05 for all simulations, indicating reliability. SUCRA scores were given to each intervention using the models.

Assessment of Inconsistency and Additional Analyses
We evaluated the consistency of the network models using the node split method with significant inconsistency defined as P < 0.05. When we identified inconsistency, we reviewed the extracted data to ensure that there were no errors and examined the potential effect modifiers of the studies. We performed further sensitivity analyses to improve the robustness of the results. We used multivariate network metaregression to analyse the remaining inconsistency. The deviance information criterion (DIC) was assessed, and we took a DIC of more than five to indicate that there were substantial differences in the models with and without the covariate [26]. The Grading of Recommendations Assessment, Development and Evaluation (GRADE) guidelines were used to assess the quality of evidence of each network [27]. Evidence was graded as high quality, moderate quality, low quality or very low quality.

Risk of Bias Across Studies
We assessed heterogeneity for each outcome using I 2 and defined I 2 of 75% and greater as substantial heterogeneity, 25%-75% as moderate heterogeneity and below 25% as low heterogeneity as defined by Higgins et al. [28]. We used a comparison-adjusted funnel plot and Egger's test to evaluate the risk of publication bias [29]. An asymmetrical funnel plot and a P-value of <0.1 on Egger's test indicated the presence of publication bias.

Study Selection
After a search and removal of duplicates, we screened 2,140 article titles and eliminated 1,685 studies. We assessed 182 articles for eligibility and included 102 articles in the systematic review [7,, while 78 articles were eligible for network meta-analysis ( Figure 1). The excluded studies and their reasons for exclusion, and the included studies and their characteristics can be found in the Appendices.
respiratory complications, 38 studies (8,803 patients) that reported renal complications, 47 studies (6,847 patients) that reported cardiac complications, 45 studies (9,335 patients) that reported wound complications and 34 studies (7,957 patients) that reported anastomotic leak were included in individual NMAs. All networks contained eight nodes. GDTFlo and GDTOD studies were the most common in most of the networks. The potential scale reduction factor (PSRF) on all the network models was less than 1.05, indicating sufficient convergence and reliability of the model. The Gelman plots and convergence plots can be obtained on request from the primary author (YX).

FIGURE 2: Network geometry plots for the different outcomes
Each yellow dot represents a node for each intervention. The grey lines represent direct comparisons between the interventions. The thickness of the grey line represents the number of studies and patients available for that direct comparison.

Risk of Bias Within Studies
The risks of bias within the individual studies are presented in the Appendices. The risk of bias summary is presented in Figure 3, and the risk of bias graph is presented in Figure 4.

FIGURE 3: Risk of bias summary
Results of risk of bias assessment representing all of the judgements in a cross-tabulation of study by entry. Green positives represent low risk of bias, yellow question marks represent unclear risk of bias and red negatives represent high risk of bias.

FIGURE 4: Risk of bias graph
The proportion of studies with each of the judgements of risk of bias. Green represents 'low risk', red represents 'high risk' and yellow represents 'unclear risk' of bias.

FIGURE 5: Comparison-adjusted funnel plot for the length of stay
A scatterplot of the standard error of each study against the mean difference of each study for the length of stay. The solid line represents the null hypothesis that the study-specific effect sizes do not differ from the respective comparison-specific pooled effect estimates. Each coloured dot represents one study.

Synthesis of Results
The results of the different NMAs for the different outcome measures are summarised below ( Figure 6 and Figure 7).

Exploration of Inconsistency and Additional Analyses
The following is a summary of the exploration of the inconsistency of the NMAs of the outcomes. The network summaries, full node-splitting analysis of inconsistency and the forest plots of the network metaregression for each outcome can be obtained upon request to the primary author (YX).
Sensitivity analyses were performed and identified that the trials of Colantonio et al. [50] and Elgendy et al. [57] contributed to the inconsistency in the comparison of complication rate and cardiac complications. The randomized controlled trial by Zatloukal et al. [128] was the source of inconsistency in renal complications analysis. The studies of Zhang et al. [130] and Lopes et al. [87] were the sources of inconsistency in respiratory complications. After removing the inconsistency, rankings remained the same or slightly altered, except in the respiratory complication analysis. RFT was found to be the most effective treatment (71%), followed by GDTLid (59%) and GDTOther (58%).
Network meta-regression was performed and showed a minimal effect of risk of bias in all analyses (Appendices).

Summary of Evidence
This systematic review encompassed 102 studies of 14,017 patients. Seventy-eight studies involving 12,100 patients were included in this NMA. Only randomised controlled trials (RCTs) of adult patients undergoing elective non-vascular abdominal surgery were included, partially fulfilling the transitivity assumption, and there were minimal effects on the risk of bias. There was significant heterogeneity in the studies reflecting the diversity of the results, and no publication bias was detected.
The effects of GDT for wound complications and anastomotic leak are the most reliable. The evidence is high to moderate quality, with no inconsistency, minimal imprecision and indirectness, and not affected by bias.
Overall, different forms of GDT were more highly ranked as interventions for the different outcomes, except for mortality.
No specific studies could be attributed to the inconsistency in LOS and mortality networks. Hence, the evidence for LOS and mortality has to be interpreted cautiously and considered to be low quality and very low quality. This could be due to low mortality in elective abdominal surgery (266 of 11,219 patients, 2.37%). Both LFT and RFT were highly ranked, which seems counterintuitive. GDTPpv has a seemingly impressive result for complication rates, renal complications, and respiratory complications but can be exaggerated by the small number of studies [3]. This highlights that GDTPpv is lacking in research and should be explored further. Pulse pressure variation and pulse variability index are promising modalities of haemodynamic monitoring that require neither expensive nor invasive equipment. The effects of GDTFlo for cardiac complications, while significant, are likely overstated.
Regarding NMA on GDT in surgical patients, our results are consistent with the findings of Zhao et al. [10]. Compared with Zhao et al. [10], we also included LFT and RFT in the current NMA with improved transitivity by only including non-vascular abdominal surgery in adults. Our study had a broader and better-defined classification system, which was based on the haemodynamic monitor used and fluid protocol in each trial. Hence, there were more studies pooled within each stratum for analysis. Also, more outcomes were measured in the current study.

Limitations
As expected, there was significant heterogeneity, a reflection of the multiple intervention arms, with a majority of studies having small sample sizes and statistically non-significant results. The lack of standardised protocols for controls made comparisons difficult as SFT differed from centre to centre. Also, several GDT studies espoused RFT principles. Future analysis may consider classifying the arms into the actual volume of fluid received instead of the haemodynamic monitor used or the authors' intentions. The populations and types of major non-vascular abdominal surgeries between different RCTs were different. The composite outcomes utilised may have introduced bias. Sensitivity analysis was used to identify inconsistency, but it may introduce selection bias by excluding articles with contradicting data or large effect sizes. Thus, we reported the original results of the network plot with the inconsistencies found.

Conclusions
In adult patients undergoing major non-vascular abdominal surgery, the use of different haemodynamic monitors and methods of GDT have different outcome-dependent effectiveness on the length of stay, mortality and different perioperative complications. GDT seems superior to either RFT or LFT in reducing perioperative complications, although the evidence is mostly low quality. The evidence was the strongest and most reliable in the outcomes of wound infections and anastomotic leak, graded as medium to high quality. Table 1 shows the excluded studies and their reasons for exclusion.