All posts by TheTraumaPro

AAST 2019 #4: Kidney Injury And The “Random Forest Model”

Brace yourselves, this one is going to be intense! I selected the next paper due to its use of an unusual modeling technique, the random forest model (RFM). What, you say, is that? Exactly!

The RFM is a relatively new method (5 years old for trauma stuff) that uses artificial intelligence (AI) to try to tease out relationships in data. It is different from its better known cousin, the neural network. The RFM tries to strike a balance of flexibility so that it can deduce rules from data sets that may not otherwise be apparent.

The authors from the trauma program at Emory in Atlanta wanted to develop a predictive model to identify factors leading to acute kidney injury in trauma patients. They assembled a small data set from 145 patients culled over a four year period. Some esoteric lab tests were collected on these patients (including serum vascular endothelial growth factor and serum monocyte chemoattractant protein-1), the sequential organ failure assess score (SOFA) was calculated, and then all was fed to the machine learning system.

The authors go into some detail about how they accomplished this work.  The main results are the sensitivity and specificity of both the RFM analysis. The RFM numbers were also converted to a regression equation and similarly examined. The area under the receiver operating characteristic curve (AUROC) was calculated for both.

Here are the factoids when using SOFA and the two biomarkers above:

  • For RFM: sensitivity .82, specificity .61, AUROC 0.74
  • For the resulting logistic regression: sens 0.77, spec 0.64, AUROC 0.72

The authors conclude that the biomarkers “may have diagnostic utility” in the early identification of patients who go on to develop AKI and that “further refinement and validation” could be helpful.

I’ll say! First, RFM is a very esoteric analysis tool, especially in the trauma world. Typically, it’s strengths are the following:

  • Requires few statistical assumptions like normal distribution
  • Allows the use of lower quality models to come up with a result
  • Shows the relative importance of each prediction feature, unlike the opacity of neural networks

The downsides?

  • It’s complicated
  • Doesn’t do well with data outside the ranges found in the dataset
  • May be difficult to interpret

But the real problem here is with the results. At this point, they are weak at best. The algorithm predicts only 4 of 5 actual cases of AKI correctly and identifies barely more than half of patients who don’t. Coin toss. A good AUROC number is better than 0.8. The ones obtained here are fair to poor at best.

I understand that this is probably a pilot study. But it seems unlikely that adding more data points will help, especially if the same input parameters are to be used in the future. I think this is an interesting exercise, but I need help seeing any future clinical applicability!

Here are my questions for the presenter and authors:

  • Why did it occur to you to try this technique? Who thought to use it? Your statisticians? What was the rationale, aside from not being able to collect any more data for the study? The origin study should be very interesting!
  • Given the lackluster results, how are you planning to “refine and validate” to make them better?
  • What future do you see for using RFM in other trauma-related studies?

I’m intrigued! Can’t wait to hear the punch lines!

Reference: Random forest model predicts acute kidney injury after trauma laparotomy. AAST Oral Abstract #11.

AAST 2019 #3: Delayed Splenectomy In Pediatric Splenic Injury

Nonoperative management of the blunt injured spleen is now routine in patients who are hemodynamically and have no evidence of other significant intra-abdominal injury.  The trauma group at the University of Arizona – Tucson scrutinized the failure rate of this procedure in children because it is not yet well established.

They reviewed 5 years of data from the National Readmission Database. This is actually a collection of software and databases maintained by the federal government that seeks to provide information on a difficult to track patient group: those readmitted to hospitals after their initial event.

Patients who had sustained an isolated spleen injury who were less than 18 years old and who had either nonoperative management (NOM), angioembolization (AE), or splenectomy were analyzed. Outcome measures included readmission rate, blood transfusion, and delayed splenectomy. Common statistical techniques were used to analyze the data.

Here are the factoids:

  • About 9500 patients were included, with an average age of 14
  • Most (77%) underwent NOM, 16% had splenectomy, and 7% had AE (no combo therapies?)
  • Significantly more patients with high grade injury (4-5) had splenectomy or AE than did the NOM patients (as would be expected)
  • A total of 6% of patients were readmitted within 6 months of their initial injury: 12% of NOM *, 8% of AE *, and 5% of those with splenectomy (* = statistically significant)
  • The NOM and AE patients were also more likely to receive blood transfusions during their first admission
  • Delayed splenectomy occurred in 15% of cases (7% NOM and 5% AE) (these numbers don’t add up, see below)
  • Statistical analysis showed that delayed splenectomy was predicted by high grade injury (of course), blood transfusion (yes), and nonoperative management (huh?)
  • In patients who were readmitted and splenectomized, it occurred after an average of 14 days for the NOM group and 58 days for AE (huh?)

The authors concluded that “one in seven children had failure of conservative management and underwent delayed splenectomy within 6 months of discharge.” They stated that NOM and AE demonstrated only a temporary benefit and that we need to be better about selecting patients for nonoperative management.

Hmm, there are several loose ends here. First, what is the quality of the study group? Was it possible to determine if these patients had been treated in a trauma center? A pediatric vs adult trauma center? We know that there are outcome disparities in spleen trauma care at different types of trauma centers. 

Next, are they really pediatric patients? Probably not, since age < 18 were included and the average age was 14. Injured spleens in pre-pubescent children behave much better than adolescents, which are more adult-like.

And what about the inherent bias in the “readmission data set?” You are looking only at patients who were readmitted! By definition, youare looking at a dataset of poorer outcomes. What if you had identified 9,500 initial patient admissions from trauma registries and then tried to find them in the readmission set. I know it’s not possible to do that, but if it were I would bet the readmission and delayed splenectomy numbers would be far, far lower.

And what about those delayed splenectomy numbers? I can’t get the percentages to match up. If 15% of the 7965 patients who didn’t have an initial splenectomy  had it done later, how does 7.2% of the 7318 NOM patients and 5.3% of the 1541 AE patients add up?

Bottom line: The usual success rate tossed around for well-selected nonoperative management is around 93% when optional adjunctive AE is part of the algorithm. That’s a 1 in 14 failure rate, and it generally occurs during the initial hospitalization. In my experience, readmissions are very rare. And that’s for adults; children tend to behave even better!

I wouldn’t consider changing my practice yet based on these findings, but the devil will probably be in the details!

Here are some questions for the presenter and authors:

  • Please provide some detail on the data set. We really need to know an age breakdown and the types of centers they were treated at, if available.
  • Discuss the potential data set bias working backwards from a database that includes only readmitted patients.
  • Please clarify the delayed splenectomy statistics to help match up the numbers.

I’m anticipating a great presentation at the meeting!

Reference: Delayed splenectomy in pediatric splenic injuries: is conservative management overused? AAST 2019 Oral abstract #8.

AAST 2019 #2: Predicting Abdominal Operation After Blunt Trauma – The RAPTOR Score

Patients with blunt abdominal injury, particularly those with seat belt signs, can be diagnostically very challenging. If the patient is stable and does not have peritonitis, CT scan is typically the first stop after the trauma resuscitation room. As many trauma professionals know, the radiographic findings can be subtle and/or not very convincing.

The trauma group at the University of Tennessee in Memphis sought to identify specific findings that might help us better identify patients that will need laparotomy. They retrospectively identified all their mesenteric injuries over a five-year period. A single blinded radiologist (is this an oxymoron or not?) reviewed all 151 patient images who underwent laparotomy, looking for predictors of bowel or mesenteric injury.  All of the predictors were then converted into a scoring system called RAPTOR (radiographic predictors of therapeutic operative intervention; kind of a stretch?). These predictors were then subjected to multivariate regression analyses to try to tease out if there were any independent predictors of injury.

Here are the factoids:

  • A total of 151 patients were identified over the 5 year period; 114 underwent laparotomy
  • Of the 114 operated patients, two thirds underwent a therapeutic laparotomy and the other third were nontherapeutic
  • There no missed injuries in the non-operated patients
  • The components of the RAPTOR score were culled from all the potential findings, and were determined to be
    • Multifocal hematoma
    • Acute arterial extravasation
    • Bowel wall hematoma
    • Bowel devascularization
    • Fecalization (of what??)
    • Free air
    • Fat pad injury (??)
  • Linear regression then showed that only three of these, extravasation, bowel devascularization, and fat pad injury to be independent predictors of injury
  • If three or more RAPTOR variables were present, then the sensitivity, specificity, and positive predictive values for injury were 67%, 85%, and 86%, and an area under the receiver operating characteristic curve (AUROC) of 0.91

The authors concluded that the RAPTOR score provided a simplified approach to detect patients who might benefit from early laparotomy and not serial abdominal exams. They go further and say it could potentially be an invaluable tool when patients don’t have clear indications for operation.

It looks like there are two things going on here at the same time. First, a new potential scoring system is being piloted. And second, a regression analysis is being used to examine the data as well. 

But first, let’s back up to the beginning. This is a retrospective study, with a relatively small size. This makes it far harder to ensure that the results will be significant, or at least meaningful. Use of a single radiologist can also be problematic, especially since many of the CT findings with this mechanism of injury are subtle. 

The reported performance of the RAPTOR score is a bit weak. The listed statistics show that it accurately identified only two thirds of those who needed an operation and 85% of those who didn’t. The AUROC for the regression is very good, though. Could a good old-fashioned serial exam scenario be better?

Bottom line: It will be interesting to hear the background on RAPTOR vs regression, and find our how the authors will use or are using these tools.

Here are my questions for the presenter and authors:

  • Why did you decide to create a scoring system that uses a set of variables that may be dependent on each other? Isn’t the regression equation better?
  • Has this information changed your practice? It seems that the two of the three regression variables are fairly obvious reasons to operate (active extravasation and devascularization). Do you really need the rest?
  • Has this study helped you decrease the non-therapeutic laparotomy rate for blunt abdominal injury?
  • And please define fecalization and fat pad injury!

I’m looking forward to hearing this presentation!

Reference: RADIOGRAPHIC PREDICTORS OF THERAPEUTIC OPERATIVE INTERVENTION AFTER BLUNT ABDOMINAL TRAUMA: THE RAPTOR SCORE. AAST 2019 Oral Paper 6.

 

AAST 2019 #1: Survival Benefit Of Pelvic REBOA

Resuscitative Endovascular Balloon Occlusion of the Aorta (REBOA) is one of the new, shiny toys in the trauma professional’s toy chest. Research papers on the topic are increasing exponentially, but human data was not even published until 2014! This is still a new device and we are trying to learn more about it.

The AAST set up an Aortic Occlusion for Resuscitation in Trauma and Acute Care (acronym is AORTA, ugh!) to help accumulate data for this not-often used technique. Hopefully, compiling comprehensive use and outcome data will speed our appreciation of the usefulness of this device.

A multi-institutional trauma group massaged the AORTA registry to examine the potential benefits of using the technique in patients with pelvic fractures leading to severe blood loss. They specifically looked for patients with the balloon inflated in Zone 3 to decrease bleeding from below the aortic bifurcation. Here’s a diagram of the zones:

The authors identified a total of 109 patients pelvic fractures with bleeding from below the bifurcation.

Here are the factoids:

  • The presenting patients arriving without CPR all had similar base deficit, lactate, and systolic BP. This shows us that the two groups are the same, but only for these three parameters. GCS was lower in the open aortic occlusion group. This could certainly contribute to a higher overall mortality in this group.
  • Overall mortality was significantly lower in the REBOA group that included those arriving with CPR in progress (35% vs 80% for open occlusion)
  • And when CPR patients were excluded, the mortality was significantly lower (33% vs 69%)
  • One in ten patients undergoing REBOA suffered vascular access complications (vascular repair required, limb ischemia, distal embolization, or amputation)
  • Complications among survivors were not different between the groups, nor were hospital or ICU lengths of stay or blood usage

The authors state that their data shows a “clear survival advantage” in those patients who undergo REBOA. Furthermore, this was accomplished without increasing systemic complications. They finally conclude that REBOA should be “strongly considered” for patients in shock due to pelvic trauma.

Not quite so fast here. There are several more factors in play than meet the eye.

First, a study that massages a REBOA database was generally constructed to see if REBOA is beneficial, especially in this time of rapid investigation. And it was performed by institutions who are using it regularly. This could introduce a significant degree of confirmation bias, since we all try to see what we already believe to be true (“REBOA is good”).

The authors are basing this “clear survival advantage” on overall mortality where only a few confounding factors have been controlled for. The GCS wild-card here is a perfect example. It could have considerably contributed to mortality in the open group, making it look bad. Who determined whether REBOA or open technique would be used, and why? This can have a major impact. What other factors might be present that are not even recorded in the database?

It is also stated that this increased survival was accomplished without increasing systemic complications. Perhaps, but that may be true of only the ones examined, or those recorded in the registry. Many may be missing. And what about the 10% incidence of limb issues in the REBOA group? This is a major problem and should not be glossed over. Although the patients that required a vascular repair were reported to do well, the others with ischemia or limb loss obviously did not.

Bottom line: Reading abstracts is like reading scientific papers, only more difficult because information is missing due to length limitations. Look at the title. Look at the conclusions. But don’t believe anything until you can understand every one of the results listed. And be sure to think about all the things that have to be left unsaid because of the size of the abstract! 

Having said all that, I still have to be careful that this doesn’t trigger my own confirmation bias. My take is that REBOA is still an investigational device. We need further comprehensive data to make sure that survival and safety are properly balanced.

Here are some questions for the presenter and authors:

  • The abstract describes the number of cases identified as 109; 84 REBOA and 25 open occlusions of the aorta. This seems to include patients undergoing CPR upon arrival, and these are excluded from some of the statistics. However, I can’t get the mortality percentages to match for the group that supposedly includes CPR patients. For example, the overall REBOA (includes CPR) mortality percentage is 35.17%. Multiplying this by 84 gives 29.5 patients. But multiplying the 33.33% mortality (CPR-excluded group) by 84 yields 28 patients. So are the 109 patients listed in the abstract the CPR-excluded group or not?
  • The open aortic occlusion group had a lower GCS. Did you look at how this might have contributed to the higher observed mortality? Although numbers are already low, is there any way to match for this to clarify the picture?
  • Do you have any information yet on longer term outcomes in the two groups? This will become very important as we come to balance raw survival with quality of life and complications.

Great abstract! I’m looking forward to the presentation, and hopefully more answers!

Reference: SURVIVAL BENEFIT FOR PELVIC TRAUMA PATIENTS UNDERGOING RESUSCITATIVE ENDOVASCULAR BALLOON OCCLUSION OF THE AORTA: RESULTS OF AAST, AORTIC OCCLUSION FOR RESUSCITATION IN TRAUMA AND ACUTE CARE SURGERY (AORTA) REGISTRY. AAST 2019 Oral Abstract #3.

Coming Up! AAST 2019 Abstract Reviews!

It’s that time of year again. The 2019 Annual Meeting of the American Association for the Surgery of Trauma (AAST) is only two weeks away! I’ll be selecting a number of interesting abstracts (oral, poster, and perhaps a few quick shots) to review here.

My focus will be on abstracts that offer new information or interesting insights into old problems. I’ll pick them apart, looking at their strengths and weaknesses. Finally, after rendering my opinion of their import, I’ll list a number of questions for the authors or presenter to consider. Who knows, they may be asked some of them at the meeting?!

Enjoy, and feel free to provide your own comments here!