Our next speaker will be Reddy Rahmani with an abstract entitled Preoperative Risk Score, Predicts 30-Day Mortality Following Emergency Surgery for Intracranial Hemorrhage. Hi, everyone. My name is Reddy. I'm a neurosurgery resident at the University of Rochester. I want to thank the AANS Trauma Section for allowing me to talk about our work on devising a risk score for predicting 30-day mortality in folks that need emergent surgery for intracranial hemorrhages and specifically subdural hematomas. So this might be a scenario that many of us in this room are familiar with. Here we have a 75-year-old gentleman who fell in his nursing home. He was intubated on arrival in the ED, was found to have a blown right pupil, was extensive posturing in the left. And as you're looking through his medical chart, you realize that he's got diabetes, dialysis dependence. He's got AFib, is on Coumadin with an INR 3.1. Now, there's been a fair amount of literature that's looked into what are the predictive factors for this gentleman in terms of his neurological exam and his imaging. However, the medical factors haven't been as scrutinized. So knowing this, we wanted to see if we can analyze a large surgical database to find the factors that would be important predictors in his outcome besides his neurological function and his imaging. So what we ended up doing was we queried NISQIP from 2006 to 2015 and we used CBT and ICD-9 codes to look for craniotomies and craniectomies for subdural hematomas. Now, we realized that this is a database study, so we wanted to be very rigorous in our methodological study of the database. So we actually ended up excluding all cases that had any variables missing. We then followed this up by further excluding the numbers by looking only for cases that were booked as surgical emergencies because that, by the definition of NISQIP, would not allow that patient to have been medically optimized prior to surgery. And then the final step to really increase the internal validity of our study was, we then split that sample into a training dataset, which we will use to make the statistical models, and then a validation dataset where we will test the statistical models that we developed. So briefly, this is just a comparison of the two datasets from that 1,200, and they're essentially equivalent. Some of the things I can draw your attention to is that a majority of the patients in the NISQIP database that had emergent subdural evacuations were independent before injury. A strong minority of them were on anticoagulants, and many of them were older in age. So this brings me to our actual statistical analysis. The first thing that we did is, it's called bootstrapping, and I can quickly go over that. Essentially, what this involves is, you take a small number of subjects from the database. You run a univariate analysis on that, and then find which factors are statistically significant. You then return those subjects to the larger database, and you take out another group of X patients, and you do this however many times. In our case, we did it 10,000 times, and then you keep track of what variables are statistically significant and how much, and you can see here that we had seven factors that were highly significant in the bootstrap analysis. We took this one step further by then performing the typical multivariate analysis that most of us are familiar with, and lo and behold, the same seven factors that were statistically significant with the bootstrapping were also significant with the multivariate analysis. And here I've also included a graphical representation of the odds ratios that we have from the multivariate analysis. So we know what factors are important, and now we want to figure out how much does each factor matter in predicting mortality in these folks. And what we did is essentially we rounded the odds ratio to a whole number. Now, that's not an exact science, but there's good support in the literature for using that as a method of assigning weight to the variable. This brings us then to the analysis of how strong our models are. So this is, if I can draw your attention to this top graph here, this is the ROC curve for the statistical model for the training dataset. And as you can see here, the area under the curve is .8. Typically in the literature, anything higher than .6 is considered a strong test. So this is an excellent test for discriminating the variables predicting outcome of mortality. And as you can also see, our device scoring system works very well with an increase in percent mortality with an increase in points. The most important part of this presentation and this whole project is this graph right here. This is now the ROC curve for the training dataset where we applied our statistical models to the naive dataset. And you can see the area under the curve is .77, indicating a very strong statistical model. Now, we have our scoring system and we wanted to make it even more applicable. So we figured that a way that this could be brought to the bedside even more is to, instead of have this range of numbers, we can dichotomize a patient to the high risk of mortality versus low risk. And what we did is we found the inflection point in our graphs and we found that a score of six or higher is statistically significant for being a high risk patient to have that outcome versus a low score. So in summary, we've essentially devised a scoring system based on preoperative medical factors that can help predict the mortality of patients based on need for emergent surgery for subdural evacuation. Some limitations, as always, even though we were very rigorous with our methodology on how to analyze the database, there's always limitations with database analysis. The scoring system may need independent validation. And, you know, the final point of thought is that these seven factors may actually end up being important not just for emergent subdural evacuations, but can actually be the case for many emergent operations. And that's something that we can follow up on. So I want to acknowledge Sam Tomlinson, who's a fantastic medical student and has worked on this project with me, and also my mentor, Dr. Bates. Thank you. I want to encourage the audience to ask some questions. We are ahead of schedule. That never happens. Are there any questions from the audience? So why did you do the bootstrap analysis? What did you learn from it? So we used essentially a statistical analysis package that Sam has actually been the gentleman who helped me with. We kind of came up with the plan of what would be the best way for us to analyze this data and then prove it on itself to do that. And so what we learned is essentially that those factors were the seven most predictive for the outcomes. Yes. Yes. Your risk factor of bleeding disorders, that underlying bleeding disorder or presence of anticoagulants? Yes. So the way that it's defined in NISQIP is an underlying bleeding disorder, something genetic, or being on anticoagulants. Either or. Yes, exactly. So there's a need to, obviously, your output is just as good as what is put into the database in terms of deciding between the two. But that would be the only classification in NISQIP that if you're on an antiplatelet medication, a low-dose aspirin, that does not count as a bleeding disorder. But your anticoagulant medications were reversed for surgery, I assume. Right. But it is still a major risk factor. Right. I have a question for you. So it is to assess outcome in traumatic brain injury, but not any traumatic brain injury parameters made your model. What brain injury parameters did you throw into your bootstrap analysis? Yes. And how do you explain that? Yes. So our goal was not to necessarily, one, the NISQIP is not a great database for taking into account a lot of the brain injury parameters. It's just simply not tracked. Even something as simple as GCS is not something that's tracked in NISQIP. Our hope was more so to look at medical factors of a patient that presents to the ED and needs emergent surgery, and then how those would affect things in the big picture of the central nervous system damage that they're sustaining, as well as other factors. But you didn't have GCS, neurological cell, bone, pupils, in your database? No. No. Those are not available. OK. Thank you. Good presentation.

Reddy Rahmani

risk score

30-day mortality

emergency surgery

subdural hematomas