Analgesic Efficacy of Adjuvant Medications in the Pediatric Caudal Block for Infraumbilical Surgery: A Network Meta-Analysis of Randomized Controlled Trials

Various adjuvants are added to local anesthetics in caudal block to improve analgesia. The comparative analgesic effectiveness and relative rankings of these adjuvants are unknown. This network meta-analysis (NMA) sought to evaluate the comparative analgesic efficacy and relative ranking of caudal adjuvants added to local anesthetics (versus local anesthetics alone) in pediatric infra-umbilical surgery. We searched the United States National Library of Medicine database (MEDLINE), PubMed, and Excerpta Medica database (Embase) for randomized controlled trials (RCTs) comparing caudal adjuvants (clonidine, dexmedetomidine, ketamine, magnesium, morphine, fentanyl, tramadol, dexamethasone, and neostigmine) among themselves, or to no adjuvant (control). We performed a frequentist NMA and employed Cochrane’s ‘Risk of Bias’ tool to evaluate study quality. We chose the duration of analgesia (defined as 'the time from caudal injection to the time of rescue analgesia') as our primary outcome. We also assessed the number of analgesic dose administrations and total dose of acetaminophen within 24 h. The duration of analgesia [87 randomized control trials (RCTs), 5285 patients] was most prolonged by neostigmine [mean difference: 513 min, (95% confidence interval, CI: 402, 625)]. Dexmedetomidine reduced the frequency of analgesic dose administrations within 24 h [29 RCTs, 1765 patients; -1.2 dose (95% CI: -1.6, -0.9)] and the total dose of acetaminophen within 24 h [18 RCTs, 1156 patients; -350 mg (95% CI: -467, -232)] the most. Among caudal adjuvants, neostigmine (moderate certainty), tramadol (low certainty), and dexmedetomidine (low certainty) prolonged the duration of analgesia the most. Dexmedetomidine also reduced the analgesic frequency and consumption more than other caudal adjuvants (moderate certainty).


Introduction And Background Introduction
A caudal epidural block is a common regional analgesic technique in pediatric surgery [1]. It is a time-tested, safe, and efficacious technique [2]. However, the duration of post-operative pain seen with much pediatric surgery (>24 h) outlasts the duration of analgesia afforded by a standard 'local-anesthetics only' caudal block (4-12 h) [3]. While continuous catheters prolong analgesic duration, such techniques are more cumbersome, require significant technical expertise [4], and may be associated with higher adverse events. Contrary to this, adding adjuvants to local anesthetic is an appealing alternative. Adjuvants can improve the block and analgesic duration [5], reduce general anesthetic [6] or local anesthetic requirements [7], allow for smoother emergence, lower incidence of emergence delirium [8], and facilitate early discharge in ambulatory surgery.
Various adjuvants have been shown to enhance caudal blocks with varying degrees of success. A multitude of clinical trials and meta-analyses have analyzed the efficacy of different adjuvants such as alpha-2 agonists (clonidine [9] and dexmedetomidine [8]), N-methyl-D-aspartate (NMDA) agonists (ketamine [10] and magnesium [11]), opioids (fentanyl, morphine, and tramadol [12]), corticosteroids (dexamethasone [13][14]), and acetylcholine esterase inhibitors (neostigmine) [12]. The European Society of Regional Anesthesia and Pain Therapy (ESRA) and the American Society of Regional Anesthesia and Pain Medicine (ASRA) joint committee practice advisory on pediatric regional anesthesia [3] provides specific recommendations on many adjuvants but given a plethora of recent studies; this advisory is likely already outdated. Furthermore, while each adjuvant is superior to the control (no adjuvant), it is difficult to ascertain the most efficacious agent (or their comparative rankings) based on clinical trials or meta-analyses alone. Network meta-analysis

Summary Measures
We extracted continuous data as mean and standard deviation (SD). When median and range were available, these estimates were derived using the method described by Hozo et al. [19] and Wan et al. [20]. We used simple imputations to impute SDs when not reported [21]. For continuous outcome, we used the weighted mean difference (WMD) with 95% confidence intervals (CI) to measure the difference in effect size between each pairwise comparison. We interpreted the potential differences in results between groups in the context of a minimal clinically important difference (MCID) of 25% of the effect size of outcomes in the control groups for each outcome. We identified this as 100 min for the analgesic duration, 0.5 doses for the number of dose administration, and 120 mg of acetaminophen for the analgesic dose. We arrived at this definition of MCID through discussion and consensus among the local intra-department clinicians. We have described our detailed statistical methods in the Section 2 of the Appendix.

Statistical Analysis
We used the R-statistical package (R Studio v 1.4.1) for frequentist statistical analysis (netmeta package [22]). We also employed frequentist methods using STATA v 14.0 (StataCorp, USA; network package [23][24]) and Bayesian methods in R Studio (BUGSnet package [25]). The details on the use of multiple packages (with reasons) are provided in the appendix. Two authors (H.S and U.S) performed the statistical analysis and checked for errors by the third (JM). We conducted a pairwise frequentist metaanalysis using the DerSimonian Laird random-effects model [26]. We considered differences statistically significant if p < 0.05 (two-sided) or when values of 0 and 1 were not included in the 95% CI for continuous and dichotomous outcomes, respectively. We used the I2 statistic to identify statistical heterogeneity [27]. We employed contrast-based parametrization [28], data augmentation, and assumed common heterogeneity variance across all pairwise comparisons. We assessed network geometry, assigning the node size that reflects the corresponding sample size and arm width that reflects the corresponding number of studies [29]. We obtained the resultant mixed (or network) estimates assuming the consistency model (i.e., heterogeneity is independent of the comparison examined) and constructed league tables of mixed estimates for each outcome. We assessed each network's global inconsistency (frequentist and Bayesian) and local inconsistency. Using the contribution matrix, we analyzed the contribution of each mixed estimate's direct vs. indirect comparisons [18]. We produced a ranking of the adjuncts for each outcome of interest using the surface under the cumulative ranking curve (SUCRA) [23], yielding a probability (percentage) of an intervention being among the best options and a mean rank. Finally, we combined results from all analgesic outcomes to ascertain the best adjuvant across all analgesic outcomes using a 'rank-heat plot' [30].

Assessment of Inconsistency
Inconsistency may invalidate the findings of an NMA. We evaluated inconsistency between the direct and indirect estimates using the global approach in both frequentists (design-by-treatment model, Higgins and co-workers [31]) and the Bayesian framework (leverage plot [25]). We also visually inspected the network forest plots to assess agreements between the consistency and inconsistency models in the frequentist method (Wald test) as well as Bayesian methods (DIC and model performance). We investigated local inconsistencies using node-splitting [32]. We planned to present results as mixed estimates if global inconsistency was not detected. We downgraded the evidence if we identified significant local inconsistencies.

Publication Bias
We evaluated statistical evidence of publication bias for each outcome for pairwise comparisons by visually inspecting Begg's funnel plot for asymmetry and conducting an Egger's regression test [33]. At the network level, publication bias was assessed using a 'comparison-adjusted' funnel plot' [34]. This depicts the difference between the study-specific effect sizes from the corresponding comparison-specific summary effect for each comparison in a network and plots this on the horizontal axis. The 'comparison-adjusted' funnel plot should be symmetric around the zero line without small-study effects.

Additional Analysis
We recognized that clinical and methodological differences between studies potentially introduce significant statistical heterogeneity. Thus, we planned to explore this heterogeneity using subgroups analysis (risk of bias and type of local anesthetic) and meta-regression analysis (local anesthetic volume and concentration; adjuvant dose). We performed such network meta-regression using a Bayesian framework (frequentist package 'netmeta' in R is unable to do so). We anticipated only a few studies to use lidocaine or epinephrine. Thus we did not study a formal analysis of the use of such agents, as it would likely lead to disconnected networks.

Grading of Recommendations
We assessed the certainty of evidence from the NMA results using the GRADE approach [35,36] using CINeMa platform and methodology [18]. Such an assessment differs from the pairwise meta-analyses in critical aspects. Six domains that affect confidence in the NMA results are within-study bias, reporting bias, indirectness, imprecision, heterogeneity, and incoherence (or inconsistency). In this way, reviewers assess the level of concerns for each relative treatment effect from NMA as giving rise to 'no concerns,' 'some concerns,' or 'major concerns' in each of the six domains. Finally, we summarized judgments across the domains into a single confidence rating ('high,' 'moderate,' 'low,' or 'very low').

Study Selection
Our search identified 1132 records, which yielded 759 records after de-duplication. Of these, we screened 252 full-text records for eligibility. Finally, we included 89 unique records in this review. This screening process is summarized in Figure 1 (PRISMA flow diagram) [16].

Risk of Bias Assessments
For the primary outcome, duration of analgesia (n=87 RCTs), we adjudged 32 RCTs at low risk of bias, 48 RCTs with some concerns, and 7 RCTs at a high risk of bias. For the number of dose administrations (n=29 RCTs), we adjudged 11 RCTs at low risk of bias, 15 RCTs with some concerns, and 3 RCTs at a high risk of bias. For the number of dose administrations (n=18 RCTs), we adjudged 8 RCTs at low risk of bias, 6 RCTs with some concerns, and 4 RCTs at a high risk of bias. Inadequate details about randomization and allocation concealment were the most common reason for downgrading the rating, followed by concerns about outcome measurement. We have summarized these results in Table 5.

Results of Pairwise Meta-Analyses
All adjuvants significantly extended the analgesic duration compared to control except magnesium and morphine. All adjuvants except dexamethasone significantly reduced the number of doses required within 24 h. All adjuvants except clonidine reduced the total dose of acetaminophen needed within 24 h. These results were associated with significant heterogeneity (I2 > 50%), perhaps due to varying concentration and dosing of local anesthetic within studies. Formal publication bias assessment was not possible as many comparisons had fewer than 10 studies. Visual inspection of funnel plots did not suggest publication bias. We have summarized these results in the Section 3 in the Appendix.

Network Geometry
We were able to assess all planned outcomes. The duration of the analgesia network constituted 10 interventions and was assessed in 87 RCTs (n=5285 patients). The most dominant nodes in this wellconnected network were control (no adjuvant) vs. dexmedetomidine (n=21 RCTs), clonidine (n=20) and ketamine (n=14). The number of dose administrations network constituted eight interventions and was assessed in 29 RCTs (n=1765 patients). The most dominant nodes in this network were control (no adjuvant) vs dexmedetomidine (n=8 RCTs), clonidine (n=5), and tramadol (n=5). The total dose of the acetaminophen network constituted ten interventions and was assessed in 18 RCTs (n=1156 patients). The most dominant nodes in this network were control (no adjuvant) vs dexmedetomidine (n=4 RCTs), ketamine (n=3), and tramadol (n=3). These characteristics are shown in Figure 2. The red circles represent interventions in each network, while a gray line connecting any work interventions represents a trial (or a trial arm in case of multi-arm studies). The total number of comparisons between any two interventions is printed as a number (in blue) on the respective gray line. Each intervention (red-circle) carries a label with its respective caudal adjuvant for each outcome. a. The network for primary outcome 'duration of analgesia' constituted 10 interventions and was assessed in 87 RCTs (n=5285 patients); b. The network for 'number of dose administrations' included eight interventions and was assessed in 29 RCTs (n=1765 patients), and c. The 'total dose of acetaminophen' network constituted ten interventions and was assessed in 18 RCTs (n=1156 patients).

Results of Network Meta-Analyses
Our analysis revealed that compared to control, neostigmine (WMD 513 min, 95% CI 402-625 min; n=9 RCTs, moderate certainty) prolonged the duration of analgesia the most, followed by tramadol (WMD 320 min, 95% CI 229-410 min; n=10 RCTs, low certainty) and dexmedetomidine (WMD 310 min, 95% CI 242-377; n=21 RCTs, low certainty). Based on an MCID of 100 min, morphine, magnesium, and fentanyl were not significantly better than control. Treatment rankings and SUCRA suggested that neostigmine was the best adjuvant, followed by tramadol and dexmedetomidine.
Compared to control, dexmedetomidine was most effective at reducing the required number of doses within 24 h (WMD -350 mg, 95% CI -467, -232 mg, n=4 RCTs, moderate certainty). While morphine also reduced this dose (WMD -373 mg, 95% CI -610, -135 mg, moderate certainty), this evidence was an indirect comparison. Based on an MCID of 120 mg for acetaminophen use, no other adjuvant was superior to control. Treatment rankings (SUCRA) suggested that dexmedetomidine was the best adjuvant, followed by morphine. These results are depicted in Figure 3 (network plots) and Figure 4 (SUCRA plots) and summarized in Table 6 (net-league tables). Each forest plot provides network estimates of included caudal adjuvants vs. control. A gray square represents the mean difference, while a black horizontal line represents the confidence interval. A vertical line represents the line of no effect. Units and values and the direction of the result are labeled below the x-axis for the respective outcome.  The x-axis shows the possible ranks, and the y-axis the ranking probabilities. Each colored line connects the estimated probability of being at a particular rank for a caudal adjuvant. The area under the cumulative rankograms is between 0 and 100%. The larger the SUCRA, the higher the treatment in the hierarchy for an outcome.  Treatments (or interventions) are reported in order of relative ranking for efficacy. Comparisons between treatments should be read from left to right. Their mean differences (and 95% confidence intervals) are in the cell in common between the column-defining treatment and the row-defining treatment. Mean differences above 0 favor the column-defining treatment for the network estimates and the row-defining treatment for the direct estimates.
We assessed all three outcomes using the rank heat-plot method described by Veroniki et al. [30]. Based on this, dexmedetomidine was judged to be the best adjuvant across all outcomes, followed by tramadol and neostigmine. Fentanyl fared worst among all adjuvants, while the control (no adjuvant) was the worstranking intervention. This is shown in Figure 5.

FIGURE 5: Rank heat plot.
Each circle ring represents a different outcome, while each section represents a different treatment or intervention. Each sector is colored according to the ranking of the treatment at the corresponding outcome. The scale consists of the transformation of three colors (red, yellow, and green) and ranges from the lowest to the highest value of the ranking statistic, such as 0%-100% according to the ranking statistics (e.g., Surface Under the Cumulative Ranking curve [SUCRA]) values. The red color corresponds to the smallest ranking statistic value (0%), values near the middle of the scale are yellow, and the green color corresponds to the highest-ranking statistic value (100%). The rank heat plot analysis suggests that dexmedetomidine is the best overall adjuvant for all three outcomes, followed by Tramadol and Neostigmine. Fentanyl was the worst adjuvant.

Inconsistency Assessment
We employed several methods to analyze inconsistency. We did not identify any evidence for global inconsistency for analgesia duration using frequentists and Bayesian methods. Exploration of local inconsistency using back-calculation methods revealed inconsistencies in clonidine vs. dexamethasone, clonidine vs. tramadol, dexmedetomidine vs. morphine, and neostigmine vs. tramadol comparisons. This was likely due to the paucity of direct trials in those comparisons. Given that there were only four comparisons among 30 for which direct evidence was unavailable, we concluded that the network for our primary outcome was consistent.
We did not identify any evidence of global inconsistency for the number of dose administrations using frequentists and Bayesian methods. Exploration of local inconsistency using back-calculation methods reassured this conclusion. We did not identify any evidence for global inconsistency using frequentists and Bayesian methods for the total dose of the acetaminophen network. Node-splitting identified inconsistency in only dexmedetomidine vs. fentanyl comparison. Overall, we were assured of consistency in the network. These results are summarized in Table 7.

Risk of Bias Across Studies
The proportion of direct evidence in each comparison loop was estimated using contribution matrices. Compared to control, network estimates for most adjuvants were predominantly informed by direct loops for all outcomes. The bias risk within each outcome's comparison loop was also assessed and used to inform certainty of evidence. Most loops were at some risk of bias, as shown in Figure 6. The comparison-adjusted funnel plot assessment did not yield any asymmetric plots, suggesting the absence of statistical evidence of publication bias. These results are shown in Figure 7.

Results of Additional Analysis
We assessed the impact of the inclusion of RCTs at high risk of bias (n=7 RCTs) using sensitivity analysis. The exclusion of these RCTs had no impact on the network estimates or the rankings of adjuvants. We also assessed the impact of volume-based dosing in caudal blocks in our studies through Bayesian network metaregression. This confirmed that our findings were robust and not affected by variations in volume-based dose in RCTs included herein. Similarly, we did not identify any impact of the variation of concentration of local anesthetic used in the included RCTs on any outcome. We could not assess the impact of the type of local anesthetic and adjuvant dosing on outcomes due to resulting network disconnections and the fact that different adjuvants are used in different doses.

Summary of Findings
Using the assessments above, we rated the certainty of the evidence for all analgesic outcomes. These results are shown in Table 8.    *NMA estimates are reported as weighted mean differences (WMDs) and 95% confidence intervals (CIs) as a frequentist model has been used. **Rank of treatment provides the comparative rankings of the treatment (best to worst) for a given outcome. The mean ranks and surface under the cumulative ranking curve (SUCRA) are also displayed. ***Indicated network meta-analysis estimates from indirect evidence only (no direct evidence available). Reasons for downgrading certainty assessment: 1 -Risk of bias; 2 -Heterogeneity; 3 -Inconsistency; 4 -Imprecision.

Summary of Evidence
While previous attempts have been made to compare different adjuvants collectively [9,12], our study is the first to perform a NMA and rank caudal adjuvants in order of their analgesic efficacy for all efficacy outcomes collectively. Based on the evidence from 89 RCTs (5442 patients), our study identified dexmedetomidine as the best caudal adjuvant across all analgesic outcomes (low to moderate evidence). On average, compared to using no adjuvant, dexmedetomidine prolonged the duration of analgesia by 310 min, reduced the number of analgesic dose administration by 1.2 doses, and reduced acetaminophen dose by 350 mg within 24 h of surgery. While other agents such as neostigmine or tramadol improve some outcomes, only dexmedetomidine consistently exceeded the pre-defined MCID thresholds for all outcomes.
Another fascinating insight from our results was that while tramadol and neostigmine prolonged the duration of analgesia (most likely by prolonging sensory block), they did not reduce the analgesic requirements. One explanation for this observation could be the lack of demonstrable synergism between epidural neostigmine [129] and systemic opioids, as opposed to epidural clonidine [130] and dexmedetomidine [131]. Similarly, epidural tramadol potentiates lidocaine-mediated sensory blocks in animal models [132]. Still, it is unknown if there is a synergism between caudal tramadol and systemic opioids. We observed that morphine and fentanyl reduced the need for acetaminophen dose despite not prolonging the analgesic duration. This likely points to the spinal and systemically mediated analgesic actions of these opioids [133] and differential spinal selectivity [134]. Even then, the evidence for morphine was predominantly indirect, while that for fentanyl was only marginally better than control.
In contrast, caudal dexmedetomidine has been shown to mediate analgesia through local and systemic mechanisms. It binds to perineural post-synaptic a2 adrenergic receptors inhibiting synaptic transmission at pre-synaptic ganglionic sites; inhibits the release of substance P by stimulating a2 adrenergic receptors in substantia-gelatinosa of the dorsal horn, and prevents norepinephrine release at the dorsal horn [135][136]. Locally induced vasoconstriction also prolongs dexmedetomidine's locally mediated perineural effects [137]. Through systemic uptake, it binds to a2 adrenergic receptors producing centrally mediated analgesia, hypotension, bradycardia, and sedation [138][139]. However, its higher affinity to subtype 2A of a2 adrenergic receptors implies that its cardiovascular effects are less pronounced than non-selective agents such as clonidine [135,140]. One beneficial impact of observed sedation is a reduced incidence of emergence delirium [8]. Given its local and systemic effects that aid analgesia, it is not surprising that our results confirm that dexmedetomidine consistently prolongs analgesia and reduces analgesic requirements.
Several meta-analyses have compared the relative efficacy and adverse effects of various adjuvants such as alpha-2 agonists (clonidine [9] and dexmedetomidine [8]), N-methyl-D-aspartate (NMDA) agonists (ketamine [10] and magnesium [11]), opioids (fentanyl, morphine, and tramadol [12]), corticosteroids (dexamethasone [13][14]), and acetylcholine esterase inhibitors (neostigmine) [12]. However, such individual pairwise meta-analyses cannot provide all adjuvants' comparative effectiveness and relative rankings. This insight can only be obtained through an appropriately conducted NMA wherein multiple adjuvants can be assessed simultaneously, and both direct and indirect comparisons inform the mixed estimates. Indeed, our review is the first to report these estimates using a robust NMA analysis and interpretation.
Using all adjuvants for neuraxial blocks (except epinephrine) remains an off-label indication. None of the included studies in our review evaluated the long-term neurological safety of caudal adjuvants. Such effects are best ascertained by examination or a delayed (two-week) follow-up questionnaire to assess deficits. Unfortunately, a pediatric population hinders a reliable neurologic assessment. While available data from animal and human studies indicate the safety of most adjuvants [141][142][143], drawing firm conclusions will likely require robust data on neurological safety. It is unlikely that a large-sized RCT would be carried out to assess this; in its absence, we will have to rely upon animal data or observational evidence [144][145]. Therefore, our findings are limited to establishing the relative efficacy of caudal adjuvants rather than safety.

Limitations and Strengths
Our NMA is subject to a few limitations. First, available RCTs involved diverse demographics and methods, including variations in age, gender, and the type of infra-inguinal surgery. We observed variations in local anesthetics' type, dose, concentration, and adjuvant doses. We mitigated this by employing a priori subgroups and meta-regression to explore heterogeneity and downgraded the evidence where appropriate. We could not assess the impact of the type of local anesthetic and adjuvant dosing on outcomes due to resulting network disconnections. Second, we observed some local inconsistencies attributed to design-by-treatment interactions (e.g., two-arm vs. three-arm trial) or a lack of an adequate number of trials. Third, some underlying biases (e.g., randomization and allocation concealment) were inherent to the source trials, leading us to downgrade the evidence strengths. Fourth, most of our studies were relatively small (n < 100), raising the possibility of small-study effects, overestimating treatment effect sizes, and inflating heterogeneity. Fifth, variations in the definitions and outcomes assessment may have contributed to heterogeneity and impacted the similarity assumption. Sixth, while we assessed publication bias at two stages (pairwise comparisons followed by the network) and found no evidence of such a bias, we cannot rule out its existence or impact on the network. Seventh, we chose not to assess the adverse effect of individual adjuvants in this review. This was due to two reasons: in general, most RCTs show a very low incidence of most adverse effects; and such low rates of complications, when taken together in a NMA framework, yield imprecise estimates that lack the required certainty to make any actionable recommendations. Eighth, we acknowledge that SUCRA and rankings can lead to misleading interpretations. Readers should form conclusions based on the certainty of evidence rather than rankings alone. Finally, we acknowledge that the use of most adjuvants used for perineural blocks remains off-label use, and their neurological safety is not well established.
Despite these limitations, our article has several strengths. This is the first successful application of network methodology to the domain of caudal block adjuvants. It is also by far the largest meta-analysis on the topic. The internal validity of this review is enhanced by restricting inclusion to homogenous studies of a caudal block using long-acting local anesthetic agents. Further methodological strengths include prospective registration, comprehensiveness of literature search, scrutiny of network validity, and appraisal of observed differences in a predefined clinically important difference. Finally, we used the risk of bias assessment tools and GRADE recommendations designed explicitly for NMAs.

Conclusions
Our results indicate that compared to control, neostigmine (moderate certainty), tramadol (low certainty), and dexmedetomidine (low certainty) are the most effective caudal adjuvants to prolong the duration of analgesia. Dexmedetomidine (moderate certainty), ketamine (low certainty), and tramadol (very low certainty) reduce the recommended analgesic dose frequency. The dose of acetaminophen needed is reduced most by dexmedetomidine (moderate certainty) and morphine (moderate certainty). Caudal adjuvants constitute an off-label use, and further research to establish their safety is needed.
There were no methodological amendments to the protocol once submitted. The only deviation from protocol was the additional use of R software to generate other graphs and plots (using netmeta, gemtc, and BUGSnet packages). Besides this, we used STATA routines for NMA and CINEMA software to assess confidence in NMAs. We found this to be easier and automated in preference to the manual method suggested by the GRADE group. Both methods follow approximately the same methodology.
Minimally clinical important differences were estimated as 25% of the average outcome estimate across the control group (mixed estimate) for each outcome. These were estimated to be:

FIGURE 8: Forest plot showing pairwise analysis for each adjuvant vs.
control for the duration of analgesia.
Assessment of publication bias (please see Figure 9).