Hypothermia in pediatric out of hospital cardiac arrest: a reanalysis
Thomas M. Austin, MD, MS and Sandra Gonzalez, MD.
I received today’s article suggesting we review it for the PAAD from 2 regular PAAD readers Dr. Mark Schreiner and Dr. Alan Klein. The article is from New England Journal of Medicine Evidence1 which to be honest is a journal that I had never heard of. To make matters worse, I didn’t understand a word of the methodology which used a Bayesian statistical perspective to reinterpret a previously published article in the NEJM.2 Having no idea of what Bayesian statistical analysis is I asked a good friend, colleague, and statistical guru, Dr. Tom Austin, Professor Of Anesthesiology; Associate Chief, Division Of Pediatric Anesthesia, University of Florida to review the article with special emphasis on how this statistical tool is used. He recruited one of his colleagues, Dr. Sandra Gonzalez to assist him. Myron Yaster MD
Original article
Harhay MO, Blette BS, Granholm A, Moler FW et al. A Bayesian Interpretation of a Pediatric Cardiac Arrest Trial (THAPCA-OH). NEJM Evid 2022; 2 (1) DOI:https://doi.org/10.1056/EVIDoa2200196
Today’s PAAD is a reinterpretation of a 2015 study2 using a Bayesian statistical framework in order to better decipher the original data and its potential impact on patient outcomes. This manner of data reanalysis using Bayesian methods is becoming more commonplace as researchers are recognizing the inherent limitations of the p-value frequentist statistics standard in medical research. In order to determine statistical significance, clinical trials designed using frequentist methods rely on rejecting a true null hypothesis by determining that the treatment effect has a low probability (typically less than 5%) of being the same as the control. Focusing on the likelihood of not having a treatment effect rather than the probability of having one (i.e., the alternative hypothesis) is one of the main criticisms of p-value statistics. Because of this methodology, vital information stored within study data can be overlooked as was the case in the original 2015 investigation.2 Bayesian reanalysis can provide a fuller probabilistic picture of potential treatment effects.
The original 2015 randomized clinical trial2 compared hypothermia (33.0oC) with normothermia (36.8oC) after out-of-hospital cardiac arrest in previously neurologically intact pediatric patients (aged 2 days to 17 years) with the primary outcome being survival with a favorable neurobehavioral outcome at 12-month follow-up. In addition, the study looked at overall survival at one year. The original study was designed to detect a 20-percentage-point absolute difference in the primary outcome between the two groups with an 85% probability of showing this difference if it actually exists (i.e., 85% statistical power). Although good neurobehavioral outcome and overall survival at 1 year were both higher in the hypothermia cohort (20 vs. 12% and 38 vs. 29%, respectively), neither reached statistical significance (p-values = 0.14 & 0.13, respectively) in the original analysis. The lack of statistical significance was most likely due to the anticipated large treatment effect.
This reanalysis differs from the original in that any hypothermia benefit (i.e., > 0% difference favoring hypothermia) was deemed to be clinically significant rather than large differences (i.e., ≥ 20% difference favoring hypothermia). Based solely on the data from the 2015 study (i.e., non-informative prior – see following paragraph), the Bayesian reanalysis determined that the probability of any benefit from hypothermia was 94% for both neurobehavioral outcome and overall survival at 1 year. In contrast, the probability of any harm with hypothermia (i.e., > 0% difference favoring normothermia) was only 6% for either outcome while the probability of severe harm with hypothermia (i.e., > 5% difference favoring normothermia) was <1% and 1% for neurobehavior outcome at 1 year and overall survival at 1 year, respectively, based on the study data alone. Thus, the potential benefit of hypothermia in pediatric patients after out-of-hospital cardiac arrest greatly outweighs the risk based only on the original 2015 study data. From the original investigation, the median absolute benefit of hypothermia was 6.8% and 9.1% for neurobehavior outcome and overall survival at 1 year, respectively.
Another advantage of Bayesian statistics is its ability to include previous knowledge into the present analysis in the form of informative priors (contrary to non-informative priors, which contain no information, making the analysis based only on the collected data). By doing so, the current study can build upon earlier empirical evidence and/or contemporary beliefs to provide a more comprehensive result for the study hypothesis at hand. Since similar studies in this pediatric population are lacking, the authors created Bayesian statistical models based on two types of informative priors: (1) downweighted (less importance than the study data) results from 3 adult studies; and (2) 9 standardized priors consisting of 3 beliefs (pessimistic, skeptical, and optimistic) each with 3 belief strengths (weak, moderate, and strong). Utilizing the latter standardized set of priors minimized bias since this practice was previously established and multiple studies with similar methodology have already been published.
For the models that incorporated informative priors, the probability of hypothermia being beneficial was greater than the probability of hypothermia being harmful for both outcomes except in the models that incorporated the most pessimistic prior beliefs (i.e., moderate and strongly pessimistic). In a large majority of all models used in the reanalysis (10 out of 13 including the non-informative prior model), the probability of hypothermia benefit was > 75% for both good neurobehavioral outcome and overall survival at 1 year. Given these results and the fact that no empirical evidence presently exists in any population that reflects the hypothetical most pessimistic prior beliefs, hypothermia provides modest benefit in pediatric patients who suffer out-of-hospital cardiac arrest with high probability.
PS from Myron: Despite decades of laboratory and human studies investigating how to improve neurological outcomes following cardiac arrest, very little has really worked, except perhaps maintaining normal blood pressure, avoiding hyperthermia and seizures, etc. This article which suggests hypothermia may work in some subgroups is controversial. Let me know what you think and I’ll post in a reader response.
References
1. Moler FW, Blette BS, Granhlom A, et al. A Bayesian interpretation of a pediatric cardiac arrest trial (THAPCA-OH). N Engl J Med Evidence. 2022;2(1):1-12.
2. Moler FW, Silverstein FS, Holubkov R, et al. Therapeutic hypothermia after out-of-hospital cardiac arrest in children. The New England journal of medicine. May 14 2015;372(20):1898-908. doi:10.1056/NEJMoa1411480