Last Halloween in a "Fifth Tuesday" post, I commented on an editorial to an article that I said I hoped would appear in a subsequent AAP Grand Rounds. It finally did this month!
Source: Laptook AR, Shankaran S, Tyson JE, et al. Effect of therapeutic hypothermia initiated after 6 hours of age on death or disability among newborns with hypoxic-ischemic encephalopathy: A randomized clinical trial. JAMA. 2017 Oct 24;318(16):1550-1560. doi: 10.1001/jama.2017.14972. See AAP Grand Rounds commentary by Dr. Jonathan Mintzer(subscription required).
My Halloween post had to do with Bayesian reasoning, a foundational concept in Evidence-Based Medicine, and how it is used to help clinicians and patients make informed healthcare choices. The article (and editorial) in question focused more on the use of Bayesian statistics in clinical research; same underlying principles, but a whole lot more math and not as easily understood.
The Laptook, et al, study is a monumental attempt to further define the utility of therapeutic hypothermia for near-term newborns with hypoxic-ischemic encephalopathy (HIE). Several prior studies have pretty much established its benefits when administered within the first 6 hours of life for such infants. The 6-hour cutoff was chosen based on animal data, which of course often don't precisely reflect human medicine. Furthermore, anecdotal and observational reports suggest that infants might benefit even if cooling is started beyond 6 hours of age. It's a great opportunity for a new study, but the problem is that it is very difficult to perform a large enough clinical trial to come up with definitive answers given the relatively small (but important) improvement in outcome. The authors randomized 168 newborns 6 to 24 hours of age to receive either conventional temperature control or therapeutic cooling for 96 hours. It took them 8 years of enrollment at 21 medical centers to accumulate that cohort, giving us some idea of the difficulties in conducting the study. They looked at outcomes of death or moderate to severe disability at ages 18 to 22 months. As expected, given the low sample size, the statistical analysis did not show a difference in those outcomes in the 2 treatment arms. However, the researchers had some tricks up their sleeves.
The study planners knew up front that they couldn't come up with a sample size large enough, in a reasonable period of time, to be able to answer this question definitively using frequentist statistics. Frequentist statistics are what we are all accustomed to in reading the medical literature; Bayesian statistical analysis is very different and relies on inferences made from prior studies to estimate the probability of a treatment effect. This study on cooling in HIE is exactly the situation where Bayesian analysis can be informative. The analysis found that cooling "resulted in a 76% probability of any reduction in death or disability using a Bayesian analysis with a neutral prior..." as well as "a 64% probability of at least 2% less death or disability." OK, now I can see most of you scratching your heads over that mouthful, but please read those statements again and stick with me.
In a world where randomized controlled trials can recruit a large enough sample size to be able to judge significant differences, frequentist and Bayesian analyses will essentially come up with the same answers; incorporating "prior information" in a formal analysis won't make much difference. However, in the small sample size dilemma, prior information is a key component in helping understand the data.
These outcomes of HIE, specifically mortality and serious disability, are extremely important. Even small differences in percentages are worthwhile, especially if the intervention itself is relatively safe. An absolute risk reduction (ARR) of a percent or two is valuable here, whereas in a less severe outcome (say the duration of fever in strep throat), we wouldn't care as much about an ARR of 1%. So, a Bayesian analysis is useful in this setting. But what's that stuff about "neutral prior?"
The researchers prespecified (i.e. before the study began) 3 assumptions (or distributions) about prior information. The neutral prior assumed that the benefits of cooling disappear completely if an infant is enrolled after 6 hours of age. They also considered an enthusiastic prior (the benefit doesn't diminish after 6 hours) and a skeptical prior (the intervention is harmful if implemented after 6 hours of age). They found a 90% probability of any reduction in death or disability with the enthusiastic prior and a 73% probability for the skeptical prior. So, all 3 assumptions suggest a benefit of cooling.
Although we can't conclude that there was a statistically significant improvement in outcome with therapeutic cooling after 6 hours of age, we can imply that there is a pretty high probability that such is true. Given the problems with performing a large enough randomized trial, this is probably the best information neonatologists can use to inform judgment. As some of you might be aware, however, over the 8-year time period of this study new information from another prospective study found increased mortality in infants cooled for 120 hours compared to 72 hours. Where that leaves the 96-hour cooling period is anyone's guess as to risk versus benefit. Time for another Bayesian study?
