So part of the reason why we started with an icebreaker is that we are going to plunge off a cliff with everybody's favorite subject, statistics and trial design. And I figured that we would get that out of the way partially because it's something that there's usually a question or two on every board. And the second reason that we're doing this is that there will be some synergy in that the last talk on Friday afternoon will involve some new guidelines that either have just come out or have come out in the last year. And some of the things that I'm going to talk about with this will fit in nicely with understanding the new guidelines. So if anybody needs to take a sip or two of coffee before I go into the biostatistics part, now is a great time to do this. But we're going to go ahead and start out with some very basic information. Again, it's usually there's a couple of questions on the boards. And in your practice, when you're looking at whether or not you should follow a particular study for your patients, it's at least useful to have some basic information. So the only disclosure that I'll give to you that I want to highlight is that I am a strong believer that random care is inferior to standardized care. And that is part of why I actually enjoy some of this stuff that I'm going to talk about. So to start out, I just want to talk about four of the things that are extraordinarily important when you design a trial. And my guess is most of you have seen this dozens of times. But when I design a clinical trial, these are four things that I will come back to. And when I present this to potential sites, I will use the PICO algorithm, which has to do with what population are you studying in. Are your patients in Duluth, Minnesota? Are your patients in Durban, South Africa? And are they patients on vasopressors who are hypotensive? Or are they not on vasopressors and hypotensive? Which population you're studying is extraordinarily important. The intervention that you give is the thing that you're doing, whether it's a drug, whether it's using bedside sonography to figure out whether you should give more fluids. But the intervention is in contrast to the comparator. The reason that I mention this starting out is that if you are doing a trial to see whether or not doing bedside sonography improves the outcomes of your septic shock patients, if all of the comparators, if all of your colleagues are not comfortable not using a sonogram, if the comparator includes some people who are getting the intervention, it may or may not change what happens. So one of the examples that probably many of you know is in the recovery trial, the comparator for COVID, it showed that steroids work for severe COVID if you're on oxygen. When they studied baricitinib in COVID, the comparator did not get steroids. So we don't know yet whether baricitinib works with steroids. Lots of people have strong opinions, but we don't know the answer. And finally, the outcome. Are you looking at something immediate, or are you looking at something long term? And just when I was making these slides, the preoxy study had just come out. So I figured I'd spend two minutes just talking about how this fits into the PICO algorithm. So the population is listed there. Population includes people who are being relatively urgently intubated. They're not taking people who need a bronchoscope to intubate. They're looking at people who required sedation and the use of a laryngoscope. The intervention was putting on a mask with some pressure and some peep. The comparator was a non-rebreather, or a bag mask. And you'll notice the comparator is not high flow nasal cannula. So if you're trying to decide based on this study whether you ought to use non-invasive ventilation or heated high flow nasal cannula, this study isn't going to answer that question. The outcome measure, whether somebody is whether or not they desaturated either during induction or two minutes after, or three minutes afterwards, two minutes afterwards. So it won't tell you whether a month later there's any difference. It's a really, really short term intervention. The only other thing I'm going to say The only other thing I want to talk about basic trial design with this is that one of the things about looking at a study is you want to make sure to see whether the patients were treated similarly. When they flip the metaphorical coin to see which group you're in, you'd like to see whether or not people who are on one side are similar to the others. And with 1,300 patients, usually there's not much difference. If there are 100 or 200 patients, sometimes there is. And you can see just looking at this that the patients were relatively similar. Again, I'm going to come back to the PICO questions or Steve and I will come back when we talk about some of the new guidelines that come back early Friday afternoon. So now here's a question. There's a new immune modulator coming out, and they want to find out, does it improve mortality? And the mortality was dropped down by 25% in those who got the study drug. The control arm had a mortality rate of 30%. The control arm event rate, sorry, the experimental arm had a mortality of 22%. So the absolute risk reduction in mortality was 8%. What's the number needed to treat to prevent one patient from dying? So this is a relatively straightforward calculation. You divide 1 by the absolute risk reduction, which here was 8%, which gives you a number needed to treat of roughly 12, which is a huge treatment effect. Probably antibiotics as compared with no antibiotics is the only thing I know of that will have maybe surgical source control might be the other thing for patients with sepsis, although there is no randomized controlled trial that shows that. But there are very few things that have a number needed to treat of 12 and 1 half. Low tidal volume ventilation compared with high tidal volume ventilation is the only critical care thing that comes to mind. And the second calculation, so number needed to treat for those of you who are going to take the boards is something that you're going to see. The relative risk is probably not something that you're going to end up seeing, but probably ought to mention that. And you just divide the mortality rate, the relative risk in the experimental arm by the control arm, and it's 0.73 for this. So a second thing that you could see is the opposite of the number needed to treat for a benefit, which is the number needed to harm. And this is occasionally a drug or a procedure will have a really, really bad side effect. And so what you're looking at is the difference in the control event rate minus the experimental event rate. And in this one, five people developed bone marrow suppression in the intervention arm, where only one developed it in the control arm. So the number needed to harm here would be 50. So if the same drug, if you need to treat 12 people to save a life and 50 people to cause bone marrow suppression, that's probably something you'd need to disclose to the family that you're more likely to live, but there is a small chance that this may cause a problem for you. Here's the thing that gives most people the fits, which is sensitivity and specificity. I have to tell you once every two years, I need to look this up to make sure that I don't switch them around. And if you've got a SARS-CoV test, and it is a test that sometimes works, so of those, 100 of them have been exposed to SARS-CoV, and 200 were tested, 70 have a positive test, and 50 who are not infected have a positive test. So this is not a great test here. So if you want to look at people at a test such as this, for the sensitivity, you look at the number of false, sorry, you look at the number of, as I've told you, I've got to look at this every couple of years to do this, those who have a positive test divided by those who actually have the disease. So you'd like to have a disease that has very few false negatives, sorry, a test that has very few false negatives. So one in which people have the disease, but the test shows negative. And here, the sensitivity is only 70%, so it's not much of a test. The converse of this is specificity. And there, you want to look at those who have false positives. And here, there are a lot of false positives. That 50 of them don't have the disease, but do have a positive test. So again, this particular test would not be useful in clinical practice. The positive predictive value, and I think this is unlikely that you will see this on the boards, but sometimes you may see it in an article, is the chance of having the disease when the test is positive. So true positives and the positive test in here, it's 58%. And the negative predictive value is showing that if you have being disease-free and having a negative test in here, the negative predictive value is 63%. I want to end with a couple of brief slides on trial design. You will rarely see a case-control study that will change practice, but they're sometimes useful for rare diseases. You will, however, see a fair number of cohort studies in which you want to take a group of patients who, for example, have been placed on ECMO for severe ARDS. And if you want to find out two years later how many of those patients have critical illness weakness, you can take all the patients who are exposed to ECMO and who survive and find out two years down the road what percentage of the survivors will have a particular outcome of interest. Most of you are familiar with a randomized controlled trial. The goal of them, again, is to randomize somebody to one of two treatments by metaphorically flipping a coin. And if the trial is large enough, you can say that likely that both of the groups, the control group and the intervention group, will be similar. More recently, because in theory giving half the patients an ineffective treatment, perhaps, since the control group in theory will be ineffective if the intervention is not harmful, there have been some newer trials that have come out over the last 10 or 15 years. And this was probably accelerated in critical care by the COVID pandemic. So there are a number of platform adaptive randomized controlled trials. I promise you, you will not see this on the boards. But you may see this in a high impact journal. And what these adaptive randomized controlled trials allow you to do is to pick people who may have a particular phenotype. So patients with ARDS who have excessive inflammation. And there's a trial that has been designed in Europe called the Panthers trial that is about to start that is putting people into specific phenotypes and randomizing them if you're in that particular domain to an agent that might suppress inflammation. Adaptive trials also allow you to test three or four different drugs and play the winner, if you will. So that if something looks like it's working after iterative modeling, you can randomize more of the people towards that drug and less of the people towards a drug that doesn't look like it's working. And I suspect you will see more of these in high impact journals. They use these a lot in cancer. And we are getting better at using them in critical care. And fortunately, that is all the statistics and trial design that I have for you. I'm glad to answer any questions either now or before. And again, I apologize for starting out your morning with something that may or may not have caused half of you to doze off. For those of you who didn't doze off, I appreciate your attention.

PICO algorithm

clinical trials

statistical concepts

trial designs

critical care research