Why This Paper

Why this paper?
I wanted to know a bit more about TTM,
I hadn’t gone over TTM2 and it seemed like a good excuse to
They take a Bayesian approach to their meta analysis
I wanted to read their rationale and see what they did
I wanted to see how they attempted to convey probability
I wanted to see if I am able to discuss basic bayesian approaches in a way that makes sense
So this talk is going to mix the stats and the paper a bit more than normal
Plan is to discuss TTM and it’s history
Discuss the paper itself and some of the stats as we go
And then a little bit more clarification on bayes factors towards the end
Starting thoughts,
I got very frustrated with this paper
I found the way they conveyed probabilities and percentages to be confusing,
That may have been just a me problem, but I found they jumped about a lot in a not consistent way
I thought the actual work was good, and after going over a few times feel like i got it,
but i don’t think it’s a nice introduction as a bayesian clinical paper

So for the archetype example of TTM, does anyone know who this is?
This is Anna Bagenholm, a swedish radiologist,
in 1999 in her late 20’s she was training as a surgeon in Norway,
and with friends from work would spend a lot of her free time skiing
She was involved in a horrible accident, where she fell headfirst through a partially frozen stream while still attached to her skies
she was stuck under 20cm thick ice in a small air pocket,
and after 40mins in the water she went into cardiac arrest (19:00)
rescue team weren’t able to retrieve her from the ice until 19:40
so bystander CPR started 40 mins post cardiac arrest,
and continued for over an hour while she was flown to hospital,
when they measured her body temperature she had a core temp of 13.7C
she went onto bypass at 21:40,
and got ROSC at 22:15
so no CPR for 40 mins,
no ROSC until 3 1/4 hrs later
in the context of one of the most profound cases of hypothermia ever
and made a full recovery with no neurological injury
and now works as a radiologist

Some of those who cared for her ended up publishing a case report in the lancet about it

Hypothermia after cardiac arrest?

Reminder of hypothermia: Mild Hypothermia (Swiss Stage 1) would be: 35-32 Moderate (Swiss Stage 2) would be 32-28, that’s when you start to get confused, and potentially stop shivering Severe (Stage 3) would be 28- (unconscious, not shivering) Stage 4, look dead (no immediate vital signs)

So there’s this potential mechanism that hypothermia is neuroprotective in cases of cardiac arrest, we know that each degree C your core temp falls, your cerebral metabolism drops by about 6% so there’s a potential that hypothermia post cardiac arrest (obviously v different to here where it was pre cardiac arrest), may prevent secondary brain injury, may reduce risk of death.

2002 - Bernard + HACA

2013 - TTM

2019 - HYPERION

2021

2021 - ESICM

2021 - TTM2

And then most recently and most significantly,

TTM 2 was published in 2021. This compared 33 degrees to normothermia with active cooling to avoid pyrexia

Powered to look for an ARR of 7.5% (or a 15% total change, allowing for harm), this failed to demonstrate significant improvements in mortality

In addition, TTM2 patients in the cooling arm had more cardiovascularly significant arrhythmias, more hypokalaemia, more vasopressors, higher lactates.

Previous studies had also suggested an increase in bleeding risk, ileus, and pneumonias.

So now we were at the point where there was no demonstrated improvement in mortality following cardiac arrest with “active cooling” in comparison to avoiding pyrexia. If you go back to papers like Bernard and HACA, their normothermia arm did allow pyrexia.

If we are wondering about non-shockable sub group, their non shockable sub group was similar size to the hyperion total, and still showed no benefit

2022 (January) Guidance

Trial History

2002	2013	2019	2021
Bernard + HACA	TTM	HYPERION	TTM2
TTM Works VS “Normothermia”	33 vs 36 doesn’t matter	33 vs 36 might matter in non-shockable	Normothermia is Fine

2022 (March) Paper

So now we get to our potentially too late meta analysis

ESICM published their systematic review and updated guidance a couple of months before this got published.

But there is still something useful in this paper, and that’s the way they approach the question

They approach the question of, does TTM benefit post cardiac arrest in a few ways. They look, like most of these studies, at both survival and neurological function

This paper looks at whether there’s evidence of any clinical benefit at all to TTM in survival, they look for evidence of a minimally significant clinical benefit, they look for evidence in the range that previous studies have previously been powered for.

They approach it in a range of ways too, from being positive about the likelihood that TTM works, to being quite skeptical

What do I want to get across here:

Population

Randomised Trials
Adult Patients
IHCA + OHCA
All Rhythms

Intervention and Control

I: TTM of 32-34C for 12hrs +
C: 36+ (Avoiding Pyrexia)

Outcome

Death
Neuro Recovery
180 Days

Papers- 3,792 Pts

Bernard
HACA
Laurent
Hachimi-Idrissi
TTM (Nielsen)
HYPERION (Lascarrou)
TTM2 (Dankiewicz)

Papers

Study	Year	Intervention	Control
Bernard	1996-1999	22/43	23/34
HACA	1996-2001	56/137	76/138
Laurent	2000-2002	15/22	11/20
Hachimi-Idrissi	1999-2002	6/14	8/14
Nielsen (TTM)	2010-2013	226/473	220/466
Lascarrou (Hyperion)	2014-2018	231/284	247/297
Dankiewicz (TTM2)	2017-2020	465/925	446/925

Hierarchical Robust Bayesian Model-Averaged Meta-Analysis

Aims

Estimate Likely Treatment Effect - RR
Null Hypothesis Testing

Relative Risk

\[ \begin{array}{l|c|c} & \text{Died} & \text{Survived} \\ \hline \text{TTM} & a & b \\ \hline \text{Normothermia} & c & d \\ \end{array} \]

\[ \text{RR} = \frac{a/(a+b)}{c/(c+d)} \]

Relative Risk - HACA

\[ \begin{array}{l|c|c} & \text{Died} & \text{Survived} \\ \hline \text{TTM} & 56 & 137 \\ \hline \text{Normothermia} & 76 & 138 \\ \end{array} \]

\[ \text{RR} = \frac{56/(193)}{76/(214)} = 0.81 \]

Priors etc

Results

Results - Bayes Factors

Death - 26

Results - Bayes Factors / Stats

\[ \frac{P(B|H_0)}{P(B|H_1)} = 26 \]

How does this differ from the last thing we saw?

Well this looks at the data from the perspective of the likelihood.

It’s asking, if a certain hypothesis was true, would you expect to see the data you saw, would the outcome you saw be a common or uncommon occurence?

They’ve asked this question twice, and then taken a ratio, how more likely is it that you would see the data you did in the null hypothesis vs, in the alternative hypothesis

What confuses me, is how we say the null hypothesis explains the data 26more times than the alternative hypothesis, but we also say the probability of the null hypothesis being true is 24% and the probability of the alt hypothesis being true is 76%

My understanding is that our bayes factor isn’t dependent on the prior beliefs, But our posterior distributions are

So my understanding is that our prior beliefs set in the model are what have dragged the posterior distribution towards the null hypothesis being false, when the data alone (as shown by the BF) suggests its true.

This brings me back to my worry about the priors they’ve used. They report using a prior distribution based on a frequentist, but the frequentist meta analyses are using some of the same earlier papers. So are they just exaggerating the potential of any benefit?

Final thoughts on this paper.

I thought this was an interesting way to approach a clinical situation with a lot of uncertainty arising from the previous trials. I thought using methods that allow us to look at the probability of multiple hypotheses was helpful and more intuitive

I found it an incredibly frustrating paper to work through and attempt to understand. I found they jumped between models and measures, and they way they conveyed things taught me more of the “how not to convey”

I’m still confused by the difference between reported bayes factors and the posterior probabilities, I assume that’s because i’ve made a mistake in my understanding, But there is a small worry that their selection of priors may be “double dipping” and using the data both for their priors and their posterior distribution.

It’s a paper that probably was made obselete in between submission and publication, as ultimately it seems to recommend the same as the most recent guidance, that it’s avoidance of pyrexia that reduces mortality rather than active cooling.

But it’s estimates of the underlying “true” effect of TTM is interesting. where they estimate that true ARR of ~ 2-4% of TTM, less than previously powered studies looking at 2-4x that size of effect, and how they highlight a traditional trial to explain this would need up to 9000 participants per arm!

Ultimately, while bayesian stats should allow us to approach questions of probablility of benefit more clearly, i don’t think this paper did that, at least for me, I’d recommend the Editorial paper “Quo Vadis” much more strongly