Podcast: Play in new window | Download

Subscribe: RSS

Date: May 10, 2023

Reference: Harhay MO, et al. A bayesian interpretation of a pediatric cardiac arrest trial (THAPCA-OH). NEJM Evidence. 2023.

Guest Skeptic: Dr. Kat Priddis is a paediatric emergency medicine consultant and trauma director at Watford General Hospital. She is part of the Don’t Forget the Bubbles team and faculty at Queen Mary University in London where she teaches part of the Paediatric Emergency Medicine MSc.

Dr. Kat Priddis

Case: You are working at the community emergency department (ED) when you receive a call from the local Emergency Medicine Service (EMS) team that they are bringing a 2-year-old boy who had a cardiac arrest at home. He had been having some upper respiratory symptoms in the previous days. Parents found him in bed that morning blue and unresponsive. They started cardiopulmonary resuscitation (CPR) until EMS arrived. 

Upon arrival at the ED, your team promptly begins high quality CPR and manages to obtain return of spontaneous circulation. As you are mentally running through your checklist for post-arrest care and preparing to transfer the patient, a team member tells you that there are potentially two hospitals in the area who may be able to accept the patient. One of the hospitals has a pediatric intensive care unit (PICU) that has the capability to perform therapeutic hypothermia but it’s further away. Which hospital should you transfer the patient to?

Background: Therapeutic hypothermia in cardiac arrest has been covered on the SGEM multiple times, all the way back to SGEM #21 and SGEM #54 and most recently in SGEM #391. Ken and Justin Morgenstern of First10EM provided a very nice summary of the history of therapeutic hypothermia that you can check out, so we won’t belabor the point. Other therapeutic hypothermia trials included Target Temperature Management or TTM trial (SGEM #82), HYPERION (SGEM #275), TTM2 (SGEM #336).

However, we have not covered the Therapeutic Hypothermia after Out-of-Hospital Cardiac Arrest (OHCA) in Children or THAPCA-OH published in the New England Journal of Medicine in 2015. We’re adding on a second paper looking at the Bayesian interpretation of the original study. 

Clinical Question: Does therapeutic hypothermia provide any benefit in neurobehavioral outcomes and survival in out-of-hospital pediatric cardiac arrest?

Original trial: Moler FW, Silverstein FS, Holubkov R, et al. Therapeutic hypothermia after out-of-hospital cardiac arrest in children. N Engl J Med. 2015.

Reference: Harhay MO, et al. A bayesian interpretation of a pediatric cardiac arrest trial(THAPCA-OH). NEJM Evidence. 2023.

  • Population: 295 pediatric patients (ages greater than two days to less than 18 years) hospitalized in PICUs at 38 children’s hospitals, who were admitted after OHCA. 
      • Excluded: Inability to randomize within 6 hours of ROSC, Glasgow Coma Scale (GCS) motor score of 5 or 6, decision to withhold aggressive treatment by clinical team, traumatic arrest
  • Intervention: Therapeutic hypothermia (target temperature 33°C). 
  • Comparison: Therapeutic normothermia (target temperature 36.8°C)
  • Outcome: 
    • Primary Outcome: Survival with good neurobehavioral outcome at 12 months. Outcome defined by Vineland Adaptive Behavior Scales (VABS-II) of 70 or higher (this is a scale from 20 to 160 with higher scores associated with better function)
    • Secondary Outcomes: Survival at 12 months and change in neurobehavioral function
    • Other Outcomes: Global cognitive score, blood product use, infection, serious arrhythmias through 7 days, 28-day mortality
  • Trial: Multinational unmasked randomized clinical trial 

Authors’ Conclusions Original Paper: “In comatose children who survive out-of-hospital cardiac arrest, therapeutic hypothermia, as compared to therapeutic normothermia, did not confer a significant benefit in survival with a good functional outcome at 1 year.”

Authors’ Conclusions Bayesian Interpretation Paper: “There is a high probability that hypothermia provides a modest benefit in neurobehavioral outcome and survival at 1 year.”

Quality Checklist for Randomized Clinical Trials:

  1. The study population included or focused on those in the emergency department. No
  2. The patients were adequately randomized. Yes
  3. The randomization process was concealed. Yes
  4. The patients were analyzed in the groups to which they were randomized. Yes
  5. The study patients were recruited consecutively (i.e. no selection bias).  No
  6. The patients in both groups were similar with respect to prognostic factors. Yes
  7. All participants (patients, clinicians, outcome assessors) were unaware of group allocation. No. 
  8. All groups were treated equally except for the intervention. Unsure. 
  9. Follow-up was complete (i.e. at least 80% for both groups). Yes
  10. All patient-important outcomes were considered. Yes
  11. The treatment effect was large enough and precise enough to be clinically significant. Unsure. 
  12. Financial conflicts of interest. Some investigators reported consulting fees from biomedical companies, but it did not seem there were obvious conflicts of interest.

Results: 1,355 patients were screened and met inclusion criteria. 475 were eligible. 295 underwent randomization with 155 in the hypothermia group and 140 in the normothermia group. Median age was two years of age and two-thirds were male. 

A higher percentage of the patients in the normothermia group had asystole as their initial rhythm (62% vs 55%), had a cardiovascular event as the cause of their arrest (13% vs 9%), and still required chest compression at time of arrival to the first hospital (73% vs 64%). 

Key Results (Original Paper): Therapeutic hypothermia did not have significant benefit on survival with good neurological function compared to normothermia.

Key Result (Bayesian Analysis): There is a high probability that therapeutic hypothermia can have some benefit in survival with good neurological function compared to normothermia.

From the Original THAPCA-OH Study

  • Primary Outcome:  Survival at one year with VABS-II scores ≥70 was not significantly different (p=0.14) between the two groups.
  • Secondary Outcomes: Survival at one year with change in VABS-II score from baseline was not significantly different between the two groups (p=0.13)

There were no statistically significant differences in blood product use, arrhythmias, infections, or all-cause mortality at 28 days.

From the Bayesian Interpretation

Probability of benefit from hypothermia was 94% for both neurobehavioral outcomes and survival at one year.

1) Selection Bias: There were a lot of exclusion criteria in the original THAPCA-OH study that may have led to selection bias. Two of the reasons why patients were excluded included: clinician judgment to withhold treatment and family not approached for consent because the doctor thought participation was not appropriate. These are highly subjective exclusion criteria

Additionally, we could not figure out why patients with GCS motor scores of 5 or 6 were excluded. Were they too “healthy?” This is especially odd because they also simultaneously excluded patients with poor function based on VABS-II score of <70 or Pediatric Overall Performance Category or Pediatric Cerebral Performance Category scores. Were they trying to find those Goldilocks patients who were “just right?”

2) Masking and Confounders: The trained interviewer who collected the VABS-II score was masked to group allocation, but they were unable to mask the treatment team to the intervention. Additionally, the clinical teams had control over all other aspects of treatment outside of the temperature goals. This lack of masking may have impacted how patients were treated based on their assigned group. 

For the patients who died while in the hospital or within 28 days, around 40% in both groups had a cause of death attributed to withdrawal of care for poor neurologic prognosis. This study was looking at the effects of hypothermia vs normothermia on survival and neurologic outcomes so excluding these patients when poor neurologic outcome is not assured may also affect the results. 

3) P-Values: There are several definitions for what a p-value is and what it is not [1-3]. The cutoff of 0.05 is arbitrary. It was proposed by Ronald Fisher back in 1925, and we’ve used it since to determine whether something is “statistically significant.” Other fields use different p-value thresholds to determine significance. Something can be not clinically significant but still be statistically significant or vice versa. Is it really that dichotomous? When a p value crosses a pre-specified significance threshold, we are tempted to say that there is no difference between the two interventions, but that is not really accurate. There can still be a difference, but it just is not statistically different.

4) Bayesian Approach and Definitions: The SGEM covered a bit of Bayesian statistics on an SGEM Xtra with Dr. Dan Lane. If we look at the numbers from the THAPCA-OH study without considering p-values, it very much looks like there is a difference between therapeutic hypothermia vs. normothermia.

The hypothermia group had 20% good neurobehavioral outcomes compared to the 12% in the normothermia group.

Bayes’ theorem is described as “revising prediction in light of relevant evidence.”

In many ways, this reflects our clinical thinking. We consider what the previous evidence has demonstrated when considering new evidence. Bayesian statistics looks at the trial in probabilistic terms. It looks at previous data (also called priors). 

  • Noninformative Priors: There is no prior existing data, meaning that we only consider the observed trial data, and every effect is possible. This helps you to pick up on small improvements. 
  • Informative Priors: Empirical (or based on meta-analysis) – The authors for this trial used the TTM, TTM2 and HYPERION trials as their priors. More on this later! 
  • Standardized/Hypothetical Priors: In this case, these were used to try to capture the full ranges of potential beliefs – important because there were NO trials on kids (only THAPCA). 

They arrived at a 94% probability of (any) benefit from hypothermia for neurobehavioral outcomes and one year survival. 

5) Why Do the Bayesian Analysis? The Bayesian analysis paper has different results because it looks at the data through a completely different lens [4].
P-values seek only to disprove the ‘null’ hypothesis (that hypothermia is not of any benefit). However this binary approach to statistics means that any actual effect is just written off. 

The Bayesian approach seeks not only to determine if there is any effect but the magnitude of the effect.

6) Creating Posterior Probability Distributions

The Bayesian framework seeks to create a posterior probability distribution. We are keeping existing knowledge (priors) → to which we add new knowledge (observed trial data – or ‘likelihood) → to create the most up-to-date idea about an intervention (Posterior Probability Distributions)

These two particular outcome measures had borderline p values which makes them ideal to apply a Bayesian framework to. 

The primary analysis looked at survival with good VABS-II scores (>70) and the group analyzed this using non-informative priors (assuming no previous knowledge about whether hypothermia was good or bad). 

Measures of density against absolute benefit difference (%) showed that (for a non-informative prior only) probability of benefit was 0.94, probability of harm was 0.06 and probability of severe harm was <0.01 for good neurobehavioral outcomes at 1 year.

7) Prioritizing Priors: They used a lot of priors in this study. How do we know which ones to use?

There were 3 from adult trials and 9 from previously created standardized frameworks. These 9 priors are hypothetical, designed to give a range.

Only the most pessimistic priors had a significant drop in benefit below 80% and even then, there was a tiny suggestion of benefit (2-4%) from hypothermia. 

The re-analysis showed low probability, even under optimistic priors, of a beneficial effect of therapeutic hypothermia as large as the postulated 20% absolute improvement.

8) Credible Intervals: This concept is broadly similar to confidence intervals, except they are not.

  • Non-informative median estimate 6.8% (Credible interval -1.9, 15.4)
  • Slight increase in mortality (2%) vs up to 15.4% improvement in benefit

This was very similar to the CI in the original trial. It does NOT matter if credible intervals cross zero because it’s a different way of looking at the data (a null or harmful effect is possible, but not very probable).

9) When Should Priors for Bayesian Analysis be Determined? Ideally for total equipoise and to remove intrinsic bias, a Bayesian trial analysis should be declared at the beginning.

Choosing the right prior will influence the results. Ideally, researchers should say in advance that they want to conduct Bayesian analysis and declare which priors they will use.

10) Down-Weighted Priors were Taken from Adult Trials Due to Lack of Pediatric Trials: The only meta-analysis info came from adult trials, most of which had different conclusions. These were ‘down-weighted’ by 50% to apply them to a pediatric population but this weighting has an influence.

Priors are designed to have an effective sample size. For the adult population trials, the team used the original sample size. Even down-weighted, the use of these priors makes a strong assumption about the relatability of adult data to kids. But the authors comment in their introduction that with approximately 7,000 out of hospital cardiac arrests annually in kids and a current trial not due for completion until 2029, we need to maximize the data that we have to help us make an informed choice. 

Using a Bayesian framework, there is a high probability that hypothermia at least moderately improves neuro-behavioral outcomes and survival at one year. However quantifying this is uncertain. 

Comment on Authors’ Conclusion Compared to SGEM Conclusion: We agree with both authors’ conclusions! The answer in most cases of evidence-based medicine is “it all depends.” 

SGEM Bottom Line: The benefit of therapeutic hypothermia on neurodevelopmental outcomes and survival for children experiencing out-of-hospital cardiac arrest is uncertain.

Case Resolution: You update the family about how their child is doing. You emphasize that you were able to achieve ROSC but there is still a long path to recovery and provide time for the family to ask questions. The family asks if there is anything more, they can do to help their child recover and you discuss that there may be a benefit with therapeutic hypothermia, but you are not certain. The family decides to transfer the child to the PICU closer to their home without the ability to perform therapeutic hypothermia so they can see him more often.

Clinical Application: Therapeutic hypothermia may offer some benefit to survival and neurodevelopmental outcomes after out-of-hospital cardiac arrest in pediatric patients. However, there is still a lot of uncertainty.

What Do I Tell the Patient/Family? Your child has suffered a cardiac arrest. We were able to restart his heart. However, there is still a lot to be done to help him recover. There are two PICUs that we can transfer him to. One PICU can do something called therapeutic hypothermia. This may offer some benefit in survival and function, but I am not certain. Let’s talk about what you value and what you would like to prioritize during this difficult time.

Remember to be skeptical of anything you learn, even if you heard it on the Skeptics’ Guide to Emergency Medicine.

Other FOAMed: 

  1. St. Emlyn’s: I’m All About the Bayes, ‘Bout the Bayes, No Treble
  2. The Bottom Line: Bayesian Statistics

References: 

  1.  Gelman, A. & Loken, E (2014). The statistical crisis in science. American Statistical Association, 70 (4), 321-327
  2. Coyne, J (2016). The Significant Problem of P-values. Scientific American. https://www.scientificamerican.com/article/the-significant-problem-of-p-values/
  3. Price R, Bethune R, Massey L. Problem with p values: why p values do not tell you if your treatment is likely to work. Postgrad Med J. 2020;96(1131):1-3.
  4. Edanz Academy. (n.d.). Frequentsit and Bayesian Statistics: What’s the difference? https://learning.edanz.com/frequentist-bayesian-statistics/
usechatgpt init success
usechatgpt init success