# Updating Evidence for Using Hypothermia in Pediatric Severe Traumatic Brain Injury: Conventional and Bayesian Meta-Analytic Perspectives*

Objective: To evaluate clinical trials of hypothermia management on outcome in pediatric patients with severe traumatic brain injury using conventional and Bayesian meta-analyses.

Data Sources: Screening of PubMed and other databases to identify randomized controlled trials of hypothermia for pediatric severe traumatic brain injury published before September 2016.

Study Selection: Four investigators assessed and reviewed randomized controlled trial data.

Data Extraction: Details of trial design, patient number, Glasgow Coma Scale score, hypothermia and control normothermia therapy, and outcome of mortality were collated.

Data Synthesis: In conventional meta-analysis, random-effects models were expressed as odds ratio (odds ratio with 95% credible-interval). Bayesian outcome probabilities were calculated as probability of odds ratio greater than or equal to 1. In seven randomized controlled trials (*n* = 472, patients 0–17 yr old), there was no difference in mortality (hypothermia vs normothermia) with pooled estimate 1.42 (credible-interval, 0.77–2.61; *p* = 0.26). Duration of hypothermia (24, 48, or 72 hr) did not show difference in mortality. (Similar results were found using poor outcome.) Bayesian analyses of randomized controlled trials ordered by time of study completed recruitment showed, after the seventh trial, chance of relative risk reduction of death by greater than 20% is 1-in-3. An optimistic belief (0.90 probability that relative risk reduction of death > 20% hypothermia vs normothermia) gives a chance of relative risk reduction of death by greater than 20% of 1-in-2.

Conclusions: Conventional meta-analysis shows the null hypothesis—no difference between hypothermia versus normothermia on mortality and poor outcome—cannot be rejected. However, Bayesian meta-analysis shows chance of relative risk reduction of death greater than 20% with hypothermia versus normothermia is 1-in-3, which may be further altered by one’s optimistic or skeptical belief about a patient.

^{1}Division of Critical Care Medicine, Department of Anesthesiology, Perioperative and Pain Medicine, Boston Children’s Hospital and Harvard Medical School, Boston, MA.

^{2}Department of Neurology, Boston Children’s Hospital and Harvard Medical School, Boston, MA.

***See also p. 388.**

Supplemental digital content is available for this article. Direct URL citations appear in the printed text and are provided in the HTML and PDF versions of this article on the journal’s website (http://journals.lww.com/pccmjournal).

The authors have disclosed that they do not have any potential conflicts of interest.

For information regarding this article, E-mail: Robert.tasker@childrens.harvard.edu

The PICU management of severe traumatic brain injury (sTBI) is challenging. There are no “level I Brain Trauma Foundation” (BTF) recommendations (^{1}). For “temperature control,” the 2012 BTF recommendations were “weak” based on “moderate” quality evidence from contradictory class II/III studies. The “level II” recommendations included “moderate hypothermia (32–33°C) beginning early after sTBI for only 24-hour duration should be avoided; and, moderate hypothermia beginning within 8 hours after sTBI for up to 48-hour duration should be considered to reduce intracranial hypertension.”

Since 2012, there has been new randomized controlled trial (RCT) evidence of hypothermia versus normothermia after sTBI in children (^{2}, ^{3}). Recent meta-analyses (^{4–7}) conclude that in pediatric patients there is no benefit of hypothermia, and it may even increase risk of mortality. In conventional “frequentist” interpretation of RCTs, the null hypothesis assesses significance using effect size, *p* value, and the CI. The *p* value is the probability of observing the effect (or more extreme one) given the null hypothesis (^{8}); it is not the probability that the null is true (^{9}). (The 2016 American Statistical Association statement is “P-values do not measure the probability that the studied hypothesis is true, or that the data were produced by random chance alone.” [^{10}]). The problem with decision-making based on *p* value is that it is misleading (^{11}): statistically significant difference is not necessarily clinically important; an insignificant difference does not indicate no difference; and studies with the similar *p* values do not provide the same level of evidence to reject the null hypothesis (^{9}, ^{12}). Therefore, decision-making is based on effect size and the CI (^{9}), but these measures do not help at the bedside. For example, the 95% CI indicates that if the RCT is repeated, then 95% of the CIs obtained will contain the true effect size. We want to know the interval that the effect size lies within, with the probability greater than or equal to 95%. Another vulnerability of this approach is the potential that patients of differing risk stratification, such as Glasgow Coma Scale (GCS) score, are lumped together.

Frequentist analysis starts with a hypothesis and uses data to judge plausibility. In contrast, Bayesian analysis is a directed decision-making approach starting with a prior belief and using outcome of an analysis to generate a “posterior” or “current” probability, which characterizes a renewed belief after observing data. That is, using both our prior belief (which is discounted in the frequentist framework) and observed data to evaluate a hypothesis (^{13}). These posterior and prior beliefs are represented by probability distributions and the observed data described by a likelihood distribution (i.e., probability of observed data conditional on the phenomenon). In hypothermia for sTBI, for example, the posterior probability of an effect of treatment could incorporate the prior probability generated from previous RCTs. Furthermore, aggregate statistical assessment can be translated into clinically relevant information using what the clinician believes from experience or interpretation of a patient’s condition. In the Bayesian framework, the “subjective” prior may be enthusiastic, noninformative, or skeptical (^{14}, ^{15}). In the past, it was impractical to implement Bayesian approaches to meta-analyses because calculating the posterior distribution requires major computation. However, using the Markov Chain Monte Carlo algorithm and computer simulation means that Bayesian analyses are now used frequently in medicine (^{16–20}).

We have used frequentist and Bayesian approaches to assess data from pediatric RCTs of hypothermia after sTBI in order to better understand current clinical recommendations (^{1}). We show that Bayesian analysis can add clinically useful information to the interpretation of clinical trials that cannot be provided by the frequentist decision framework.

## METHODS

### Search Strategy

A systematic search of English language literature using PubMed (2000 to 31 August 2016) search terms “hypothermia,” “traumatic brain injury,” “head injury,” and “intracranial pressure.” RCTs in pediatric patients were identified. We hand-searched reference lists of four meta-analyses of therapeutic hypothermia in pediatric sTBI (^{4–7}). We looked at the catalog of RCTs in the PICU population at www.picutrials.net. Last, we checked a federated literature search engine (ACCESSSS, see: http://plus.mcmaster.ca/ACCESSSS/). The abstracts were analyzed for relevance and eligible articles were assessed. Reference lists from RCTs were also checked for further resources.

### Study Selection Criteria, Quality Assessment and Data Extraction

The minimum inclusion for meta-analysis was use of induced systemic hypothermia for greater than or equal to 12 hours and survival at PICU discharge. Quality of evidence was assessed using the Grading of Recommendations Assessment, Development, and Evaluation system (^{21}). Data extraction included trial design, patient number, GCS score; hypothermia and control normothermia therapy; and outcomes of mortality and poor/unfavorable outcome. The division between poor/unfavorable and favorable outcome was defined using the Glasgow Outcome Scale (GOS) or the Pediatric Cerebral Performance Category (PCPC) (^{22}, ^{23}). In reports that used GOS, patients with scores 3 (severe outcome), 4 (vegetative state), or 5 (death) were grouped as poor/unfavorable outcome. In reports that used the PCPC, patients with scores 4 (severe disability), 5 (vegetative state), or 6 (death) were grouped as poor/unfavorable outcome.

### Statistical Analysis

In the conventional meta-analysis, we used risk ratio (RR) of death (or poor outcomes) as the effect measure to compare therapy (^{24}). “R-Studio software” (R Studio, Inc., Boston, MA; https://www.rstudio.com/) was used to model data and perform meta-analysis with “meta-package” (https://cran.r-project.org/web/packages/meta/index.html). The random-effects model was used to account for heterogeneity of populations, where the number of deaths and sample size of each trial was used to compute effect size and CI using the inverse variance method and the DerSimonian-Laird heterogeneity estimator (^{25}). A Forest plot was performed using meta-package. Heterogeneity was investigated by comparing relative risks of subgroups defined by hypothermia duration.

In the cumulative Bayesian random-effects meta-analysis, we used log (RR) in R-Studio software, where we assumed Gaussian distribution (^{26}) (rjags package, https://cran.r-project.org/web/packages/rjags/index.html). JAGS software (Just Another Gibbs Sample v4.2.0; http://mcmc-jags.sourceforge.net/) was used to sample the posterior density. A noninformative prior was used for both mean and variance of the log (RR) (^{27}). At each time point, this prior and the studies up to that time were used in the meta-analysis (^{14}, ^{15}, ^{28}). Then, the posterior probability of relative risk reduction (RRR) being greater than 0% and 20% was calculated after each study.

Last, since Bayesian analysis can incorporate prior belief, we have investigated the impact of prior skeptical and optimistic beliefs with two informative priors for mean log (RR), that is, probabilities of 0.01 or 0.90, respectively. These probabilities are commonly used in Bayesian clinical decision-making studies (^{28–33}). We have also calculated the posterior probability of relative risk increase (RRI) in mortality greater than 20%.

## RESULTS

The literature search is illustrated in **Figure S1** (Supplemental Digital Content 1, http://links.lww.com/PCC/A389). Thirty-five clinical trials were hand-searched to assess suitability for inclusion. At first, nine RCTs were reviewed (^{2}, ^{3}, ^{34–39}), and these studies were compared with RCTs in recent meta-analyses (^{4–7}). None of the studies in previous meta-analyses were missed. However, we found that two reports used subsets of previously published RCTs: Bourdages et al (^{34}) reported cases from Hutchison et al (^{38}) and Salonia et al (^{35}) reported cases from Adelson et al (^{37}). The authors reached consensus on the quality of the remaining RCTs (Table 1). The overall quality was low: two trials were high (^{3}, ^{38}), and in the other trials one was moderate (^{36}), and four were low (^{2}, ^{37}, ^{39}).

There were outlier studies (Table 2). Mean time to starting cooling after injury was 4.2–6.8 hours in six studies; in contrast, in the study by Adelson et al (^{37}) mean time was 15 hours. Site of temperature measurement differed across studies: three used esophageal temperature (^{2}, ^{36}, ^{38}); two, rectal temperature (^{37}); one, intracranial temperature (^{39}); and one, either intracranial or rectal temperature (^{3}). Intended temperature in the normothermia group was 36–37.5°C, but in one study that used exclusively intracranial temperature, the range was 37.5–38.5°C (^{39}), which is equivalent to pyrexia. Hypothermia ranged 32–34°C but the duration of temperature management varied: 24 hours (^{38}), 48 hours (^{3}, ^{36}, ^{37}), or 72 hours (^{2}, ^{39}). Rewarming after hypothermia was given in six studies, and varied: short, 12–18 hours (^{36–38}); intermediate duration, 16–35 hours (^{2}); or prolonged, 42–54 hours (^{3}).

### Classic Meta-Analysis

Statistical heterogeneity was low (*p* = 0.27; *I*^{2} = 20.4%). None of the trials found benefit of hypothermia (Fig. 1). The funnel plot of effect of hypothermia versus normothermia on mortality did not indicate publication bias (Fig. 2). Using RR of death in hypothermia versus normothermia, the pooled estimate with a random-effects model is 1.42 (95% CI, 0.77–2.61; *p* = 0.26). Table 3 shows the changes in the overall estimates when conducting the sensitivity analysis. That is, after removing one study at a time and rerunning the model, the estimates changed little—which implies the results are statistically reliable.

Figure 3 shows the Forest plot of hypothermia versus normothermia with studies grouped by duration of hypothermia (24, 48, and 72 hr). The statistical heterogeneity of these subgroups was low (24 hr, no analysis since only one study; 48 hr, *p* = 0.16, *I*^{2} = 42.2%; 72 hr, *p* = 0.20, *I*^{2} = 39.1%). The results showed no difference in hypothermia versus normothermia. Effect sizes and 95% CIs were 24 hours, RR = 1.78 (95% CI, 0.97–3.28; *p* = 0.06); 48 hours, RR = 1.29 (95% CI, 0.42–3.95; *p* = 0.65); and 72 hours, RR = 1.18 (95% CI, 0.16–8.84; *p* = 0.87). Duration of hypothermia did not show a difference in the magnitude of the association with mortality (*p* = 0.84). These same analyses were repeated using poor outcome and were no different to the results of mortality (**Fig. S2**, Supplemental Digital Content 1, http://links.lww.com/PCC/A389).

### Bayesian Meta-Analysis

After ordering the studies by month/yr of recruitment closure, we carried out a Bayesian cumulative meta-analysis using log (RR) with a noninformative prior (Fig. 4). The figure shows that after the seventh RCT (^{2}), the current probability of reducing mortality with hypothermia compared with normothermia is 0.40 (with RR < 1 or RRR > 0%). The current probability of RRR of death greater than 20%, with hypothermia rather than normothermia, is 0.28 (close to 1-in-3, see RRR > 20% curve in Fig. 4).

Figures 5 and **S3** (Supplemental Digital Content 1, http://links.lww.com/PCC/A389) show the effect of an a priori belief on current probability of RRR of death greater than 20%. Based on the seven RCTs, a skeptical belief (0.01 probability that RRR of death > 20% when using hypothermia rather than normothermia) gives a current probability of 0.16 (1-in-6) of RRR of death greater than 20% on theoretically using hypothermia rather than normothermia. However, the probability of RRI of death greater than 20% is 0.61 (3-in-5) if selecting hypothermia rather than normothermia. An optimistic belief (0.90 probability that RRR of death > 20%) gives a current probability of 0.50 (1-in-2) of RRR of death greater than 20% on theoretically using hypothermia rather than normothermia. The probability of RRI of death greater than 20% is 0.28 (1-in-3) if choosing hypothermia rather than normothermia. The Bayesian analysis for poor outcome provides similar findings and is given in **Figure S4** (Supplemental Digital Content 1, http://links.lww.com/PCC/A389).

## DISCUSSION

We have evaluated data from seven pediatric RCTs of hypothermia versus normothermia after sTBI using both conventional and Bayesian meta-analyses in order to better understand the basis for clinical recommendations.

The updated conventional meta-analysis of hypothermia in sTBI in pediatric patients confirms that the null hypothesis—no difference between hypothermia versus normothermia on mortality and poor/unfavorable outcome—cannot be rejected. However, we should be circumspect about the use of hypothermia for two main reasons. First, the overall quality of evidence is low. Second, lumping together these RCTs is questionable. There is heterogeneity with patients of differing risk stratification (range in GCS score) across the studies. Studies also used different body sites for temperature measurement, as well as different durations of hypothermia and rewarming. And, there were two obvious outliers: one study had considerably later timing (mean 15 hr vs mean 4.2–6.8 hr in the other studies) for starting cooling (^{37}) and one study used a pyrexial target temperature in the normothermia group (^{39}).

In regard to clinical decision-making, a previous conventional meta-analysis of five RCTs went so far as to say “because of safety concerns with hypothermia, we do not recommend further RCTs of the intervention in children with sTBI” (^{5}). This conclusion was based on an overall estimate showing hypothermia could increase mortality and poor/unfavorable outcome with sTBI but without statistical significance. We too have a similar finding with two more pediatric RCTs: the RR for death is 1.42 (95% CI, 0.77–2.61; *p* = 0.26). So, would, should, or could we use hypothermia in our sTBI patients? Our response to this quandary was to reframe the meta-analysis using a Bayesian approach (^{11}, ^{13–20}).

The Bayesian approach considers all sources of preexisting knowledge admissible for analysis. Both previous RCT results and different opinions are used to facilitate informed decisions likely to be meaningful to clinicians and guideline developers. First, in regard to information from RCTs, we have calculated the probability of a hypothesis (i.e., RRR of death > 20% with hypothermia rather than normothermia) given information from each trial in historical sequence. After the seventh pediatric RCT, the current probability of RRR of death greater than 20% is close to 1-in-3. The 20% threshold in RRR for mortality is used since a new therapy for critical illness is likely to be accepted by clinicians and regulators at this level (^{15}). Second, in regard to the more subjective aspect of clinician’s prior skepticism or optimism (i.e., beliefs), we have used the Bayesian approach to incorporate a range of views in assessing the effect of hypothermia in a particular clinical setting. For example, given a current probability for RRR of death greater than 20% of 1-in-3, a skeptical (0.01 probability that RRR of death > 20%) belief in the effect of hypothermia treatment on outcome changes the probability from 1-in-3 to 1-in-6, that is, less chance of RRR death greater than 20%. If we also consider this change in probability along with a current probability of RRI in death greater than 20% of 3-in-5 by being skeptical about effectiveness of hypothermia, you would have to conclude that hypothermia should not be used. That is, in comparison with normothermia, hypothermia has less chance of RRR death greater than 20% and more than even chance of RRI in death greater than 20%. In contrast, an optimistic (0.90 probability that RRR of death > 20%) belief in the effect of hypothermia treatment on outcome changes the probability of 1-in-3 to 1-in-2, that is, even chance in RRR death greater than 20%. In addition, the current probability of RRI in death greater than 20% of 1-in-3 means a less than even chance of RRI in death greater than 20%. Under these circumstances, the clinician with an optimistic belief about the effect of hypothermia in a particular setting may consider that the choice of hypothermia is a valid therapeutic option. In fact, this observation adds support to the recent proposal by Lazaridis and Robertson (^{40}) that “the benefit-risk ratio of HT has not been shown to be unfavorable.”

The different degrees of clinical belief that are formally translated into mathematical calculations in Bayesian meta-analysis warrant further discussion. We do not fully understand what clinical features lead treating physicians to have skeptical or optimistic beliefs about effectiveness of hypothermia in a particular pediatric patient with sTBI. Our specialty has not fully codified “optimism” and “doubt” about particular outcomes in sTBI in relation to use of hypothermia management. In contrast, investigators of hypothermia for out-of-hospital cardiac arrest (^{41}) and neonatal hypoxic-ischemic encephalopathy (^{33}) have better characterized subgroups that are less likely to respond to the intervention. Now, too, there is a systematic review of methods to elicit clinical beliefs from expert and nonexpert clinicians for Bayesian priors (^{18}) that could also be applied to the design of future pediatric critical care studies (^{14}, ^{42}, ^{43}).

## CONCLUSIONS

In our updated conventional meta-analysis of seven RCTs, we show that the null hypothesis—no difference between hypothermia versus normothermia on mortality and poor/unfavorable outcome in pediatric severe TBI—cannot be rejected. We also found no influence on outcome when comparing duration of hypothermia (24, 48, and 72 hr). These findings have an impact on the BTF 2012 recommendations: we cannot conclude that hypothermia for only 24 hours “should be avoided,” and we cannot conclude that hypothermia for 48 hours “should be considered.” This lack of guidance means that clinicians are left making the decision, and we know that some are choosing to use hypothermia (^{40}, ^{44}).

The Bayesian meta-analysis provides a framework for decision-making in circumstances of uncertainty (^{45}). By taking the cumulative probability of RRR of death greater than 20% and chance of RRI of death greater than 20% from seven RCTs and including one’s prior degree of skepticism or optimism about treatment effectiveness, one may be dissuaded or persuaded about treatment in a particular clinical setting. Therefore, more work is needed to understand clinical skepticism and optimism about treatment effect. Only in this way we will know how to deal with the results of seven RCTs on hypothermia in sTBI and be able to decide when hypothermia should be “avoided” and when it should be “considered” in pediatric management.

## REFERENCES

*p*-values: Context, process, and purpose. Am Stat 2016; 70:129–133. Available at: http://dx.doi.org/10.1080/00031305.2016.1154108

Bayesian analysis; meta-analysis; therapeutic hypothermia; traumatic brain injury