TXA has been the darling drug of emergency medicine and critical care since the publication of the CRASH 2 trial back in 2010. CRASH 2 demonstrated a 1.5% decrease in mortality when TXA was given to trauma patients with significant bleeding. (CRASH 2 2010) These results led to the widespread adoption of TXA in trauma protocols, and significant interest in the use of TXA in many other conditions. However, it is important to remember that we cannot be certain about the benefit of TXA after a single RCT, and there were numerous reasons to be skeptical of the CRASH 2 results. The next big TXA trial was TICH-2, which looked at TXA in patients with nontraumatic intracerebral hemorrhage, and showed no benefit. (Sprigg 2018) This was followed by the WOMAN trial, looking at TXA in postpartum hemorrhage, which (despite great hype) was also a negative trial. (WOMAN 2017) We now have the results of the much anticipated CRASH 3 trial, looking at TXA in traumatic head injury. Much like WOMAN, this trial has been released with incredible fanfare, despite rather unimpressive results.
The CRASH 3 Trial Collaborators. Effects of tranexamic acid on death, disability, vascular occlusive events and other morbidities in patients with acute traumatic brain injury (CRASH 3): a randomised, placebo-controlled trial. Lancet (London, England). 2019; PMID: 31623894 [free full text] Cincaltrials.gov: NCT01402882
This is a pragmatic, international, multi-centre, randomized, placebo controlled trial.
Adults with traumatic brain injury and a GCS less than 13 or any intracranial bleed on head CT, excluding patients with major extracranial bleeding. The patients had to be randomized within 3 hours of their injury (this was changed partway through the trial from 8 hours based on data that TXA is unlikely to be effective if given after 3 hours).
- They did add a “fundamental eligibility criterion” that the responsible physician was uncertain about the appropriateness of TXA therapy. Given that this is the first trial of TXA in this population, I don’t see how anyone could have been anything but uncertain, unless you really value opinion over science. Therefore, I don’t see the value of this additional criteria, but it does create potential for selection bias. Unfortunately, they don’t tell us how many patients were screened but not included in the trial, so it is hard to assess the impact of this decision.
Tranexamic acid: 1 gram IV over 10 minutes and 1 gram infused over 8 hours.
The primary outcome listed in the manuscript is head injury-related death in hospital within 28 days of injury in patients randomly assigned within 3 h of injury. (This was changed partway through to only include patients randomized by 3 hours.) This primary outcome does not match the primary outcome listed on clinical trials.gov or the ISRCTN registry. In both of those registries, all cause mortality is listed as the primary outcome.
They randomized 12,737 patients, 9,202 of whom were randomized within 3 hours (so they didn’t meet their calculated sample size of 10,000). There were 2,560 deaths overall (20.1%).
There was no statistical difference in the primary outcome of head injury related death (18.5% with TXA and 19.8% with placebo; RR 0·94; 95% CI 0·86–1·02).
As they collected data, they noticed more patients with GCS 3 or bilateral unreactive pupils than they had anticipated. They therefore added a sensitivity analysis excluding those patients, which showed the same absolute decrease, but just barely missed out on being statistically significant (12.5% with TXA and 14.0% with placebo; RR 0.89; 95% CI 0.8-1.0).
Mortality was unchanged (RR 0.96; 95% CI 0.89-1.04).
Disability was completely unchanged based on 2 different measures.
There was also no difference between the groups in terms of adverse events or vaso-occlusive events.
Thus far, the results are pretty negative, but there was a subgroup that everyone has been talking about. Looking specifically at patients with “mild to moderate head injury” (GCS 9-15), they note a reduction in “head injury related death” (RR 0.78; 95% CI 0.64-0.95). Unfortunately, they don’t tell us about the all cause mortality for this subgroup.
There is one other really important subgroup: TXA increased death from all other causes, balancing out the possible reduction in “head injury related death”, although the result wasn’t statistically significant (RR 1.31; 95% CI 0.93 – 1.85).
Overall, this is a great trial. It took a tremendous effort to get this done, and the authors deserve a ton of credit. It was huge (although unfortunately underpowered because they changed their inclusion criteria partway through) and had very little loss to follow-up. It was appropriately randomized and allocation was concealed. They specifically looked for harms.
CRASH 3 was also pretty clearly a negative trial. Their primary outcome was negative, but more importantly there was no change in the really important outcomes of mortality or disability. Although it may be underpowered, the lack of a mortality benefit in a 9-12,000 person trial is pretty telling.
Unfortunately, instead of presenting this trial for what it is (a negative trial, with some potentially hopeful data points worthy of future research and maybe even clinical consideration), this trial has been promoted with a degree of spin and hype that leaves science far behind. The spin starts in the paper itself, where claims of a significant benefit are made in the abstract and first sentence of the discussion section, despite the fact that the primary outcome of the trial was negative. In this respect, I think the Lancet failed miserably in both editing and peer review (much like it did with the WOMAN trial.) Things got much worse as this trial was released to the public with infographics claiming that “TXA could save 1 in 5 patients” and an official video claiming TXA “could save 10s of thousands of lives.”
Those are pretty bold claims for a study that demonstrated no difference in mortality.
Were the patients too sick?
So we better talk about the subgroups that have got everyone so excited.
There is a sweet spot in medicine (and in clinical trials) where patients are just sick enough to benefit from a therapy. Too sick, and they will die whether or not we treat them. Too well, and they will live, with or without our interventions. Only patients in the Goldilocks zone can benefit from our medical interventions.
This is an incredibly important consideration when looking at negative trials. If almost everyone enrolled has a dismal outcome, you won’t be able to demonstrate a benefit, even if the medication might be helping some patients. (Think about cardiac arrest research, for example.) On the other hand, if almost everyone survives, it can again be difficult to show a benefit, even if it really exists among the sickest patients. (It is possible that this is what happened in the WOMAN trial.) We need to find the right subset of patients in order for our interventions to work.
In CRASH 3, they ultimately enrolled a lot of patients with very severe head injuries. If all of these patients were destined to have bad outcomes, it would have been harder to demonstrate a benefit from TXA, even if it existed for healthier patients. For that reason, partway through the trial they decided to add a sensitivity analysis that excluded patients with a GCS of 3 or with bilateral unreactive pupils. It still didn’t quite reach statistical significance, but it was close. (Really, all that this type of sensitivity analysis does is make the sample size of the trial smaller. If you really didn’t think TXA could help these patients, it would have made more sense to exclude them from the outset. Instead, they just end up negating the value of their sample size calculation.)
Although the sensitivity analysis wasn’t quite statistically significant, they performed another subgroup analysis (this one truly pre-planned) focusing on patients with a GCS over 8, and they did find a statistically significant result.
In general, subgroups aren’t a very reliable source of data for clinical decisions. This is especially the case when they are not specified before running the trial, because it gives the researchers too many options to test (degrees of freedom), which almost guarantee the finding of a statistically significant result. However, even when the subgroups were prespecified, as they were in CRASH 3, they don’t usually pan out in future research. (Wallach 2017) That is why, in general, we consider subgroups and secondary analyses to be “hypothesis generating” for future research.
Putting aside the incredible hype seen in the lay media, I think that almost all clinicians agree that the CRASH 3 results warrant further study. We need follow-up RCTs to determine if benefits seen in these subgroups are real. The question is what to do in the meantime.
A number of incredibly smart people have suggested changing clinical practice and giving TXA to patients with isolated head trauma and a GCS above 8 based on the subgroups in CRASH 3. (See the Other FOAMed section below.) When making clinical decisions, I think that it is important to use all data available. Although we often talk about trials as being “positive” or “negative”, large trials like this provide us with so much more information, and it would be silly to ignore it. Therefore, if CRASH 3 had shown a mortality benefit in a subgroup of patients, I think it would be completely reasonable to give TXA to that subgroup while awaiting further data.
But, CRASH 3 did not demonstrate a mortality benefit, even in the subgroup of mild to moderate head injury. (At least, that we know of. The authors do not provide us with the necessary information to be sure.)
November 1, 2019: It turns out that there is a decrease in all cause mortality in the subgroup of patients with mild to moderate TBI. It was not included in the published manuscript, but based on private conversation with the author, all cause mortality was 6.9% in the TXA group and 8.3% in the placebo group (ARR 1.4%; RR 0.83; 95% CI 0.69-0.99; fragility index 2). As I stated above, that changes my conclusions.
November 17,2019: It has been pointed out to me that in the pre-published trial protocol, the CRASH 3 researchers say: “We do not expect the effect of tranexamic acid to vary substantially by severity of TBI and unless there is strong evidence against the null hypothesis of homogeneity of effects (i.e. p<0.001) the overall relative risk will be considered the most reliable guide to the approximate treatment effect in all patients.” (Roberts 2018) The reported p value for heterogeneity is only 0.03. In other words, according to their own protocol, they were supposed to ignore this subgroup, and use the main primary outcome as “the most reliable guide”.
Disease specific mortality rears its ugly head again
Mortality was not changed in CRASH 3. The difference in the subgroup, and the borderline difference in the sensitivity analysis, were both in “head injury related death”. That is not a real outcome.
As I have discussed in the past, disease specific mortality is a confusing and generally misleading concept. In general, it is based entirely on guesswork. (Most patients don’t receive autopsies, and so the cause of death is whatever the treating physician writes on the death certificate.) That guesswork might be good enough for every day accounting, but it doesn’t make sense in the context of testing an experimental agent, where unexpected and unusual causes of death may occur.
More importantly, if overall mortality is unchanged, disease specific mortality is nothing more than a futile exercise in accounting. Patients care whether or not they die. They don’t care what we write on their death certificates.
As an extreme example of the fallibility of the concept, imagine treating these patients with cyanide. Death due to head injury would drop by 100% (because all patients would die instead of cyanide poisoning), but patients wouldn’t be helped. Most real world examples won’t involve agents that increase overall mortality, but just shifting the cause of death from one category to another without changing all cause mortality is an unhelpful exercise.
Disease specific mortality is a surrogate outcome. It is used when it would be impossible to run a trial large enough to show a change in all cause mortality. However, CRASH 3 was a very large trial. The registered primary outcome was all cause mortality. The sample size calculation was based on all cause mortality. They collected all cause mortality data. In that context, there is no reason at all to use the error-prone, non-patient oriented, disease specific mortality as a surrogate outcome.
In this trial all cause mortality was unchanged. Although people have been obsessed with the decrease in “head injury related death” (in a subgroup), they have ignored the opposite end of the spectrum. TXA increased non-head injury related death by 31% (RR 1.31; 95% CI 0.93-1.85). Although the result is not statistically significant, it is a pretty good indication that TXA is just resulting in a different diagnosis being written on the death certificate, rather than providing any actual benefit.
Although I am always skeptical of conclusions based on subgroups, that is not my major concern with CRASH 3. The big problem with this trial is that the outcome in that subgroup is the faulty, non-patient oriented, disease specific mortality. If the subgroup had shown changed in all cause mortality, my conclusion would be different. (We would still need confirmatory studies, but I would suggest using TXA while awaiting those studies).
Is TXA harm free?
One of the major reasons that people are suggesting that we change our practice based on the subgroups of CRASH 3 is that TXA is perfectly safe. “If there is even a chance it could help, why not give it? It doesn’t hurt.”
It is true, based on some pretty large RCTs, we haven’t seen obvious harms from TXA. However, all medications cause harm. If TXA is having a physiologic effect, there are going to be side effects. Unless TXA is the only side effect free medication in existence, there will be some harms. So why aren’t we seeing harms? One possibility is that TXA is completely inert. We aren’t seeing harms because intravenous TXA has essentially no physiologic effect (meaning neither harm nor benefit). Another possibility is that there is harm, but we are missing it. (They didn’t systematically screen for clots in CRASH 3.) Finally, it is possible that this data is actually showing us harm, but we are just ignoring it.
I think it is important to consider the harms that are obvious in the TXA data. If we are going to focus on subgroups with potential benefit, it is pretty important to also consider their counterparts. If TXA provides no benefit overall, but a subgroup benefits, that means there must be another subgroup that is harmed to balance out the data. Looking at the results, that is exactly what we see. In CRASH 3, the authors are focused on the benefit seen in “head injury related death”, but what about the patients who don’t die from their head injury? As I already mentioned, TXA appears to increase mortality from other causes (RR 1.31).
Similarly, people have looked at the subgroups who were treated earlier in CRASH 2 and WOMAN, and determined that TXA is excellent if given in less than 3 hours. However, the same data indicates an increase in mortality if TXA is given later than 3 hours.
It isn’t clear to me whether these subgroups are real. However, you cannot simultaneously claim that TXA works in certain subgroups and is also perfectly safe. If the subgroups are true, then there are also subgroups in which TXA increases mortality. That is a pretty big harm that cannot be ignored.
There are also other harms to be considered. I have no problem giving a rapid dose of TXA, but to follow this protocol I need to set up an 8 hour infusion. If, as this data suggests, TXA isn’t providing any real benefit, that infusion may cause harm. It might occupy an IV that is better used for a different intervention. It may occupy nursing time or attention better focused elsewhere. It might use the last IV pump in the department (they are always in short supply), delaying care for other critically ill patients. Those are real concerns for an intervention that has no proven benefit.
November 1, 2019: My conclusions from this paper changed somewhat based on the unpublished data (see box above) that says that there is a decrease in all cause mortality in the subgroup of patients with GCS 9-15, especially when combined with the subgroup in CRASH 2. I still caution about the problems with subgroups. They are hypothesis generating for a reason – most don’t pan out in future research. I also caution that the CRASH3 protocol specifically says that this subgroup should not be considered as reliable as the primary outcome. However, this is a change in a very important outcome. I still don’t think guidelines should strongly recommend this treatment. More research is clearly needed to confirm this finding. It is not fair to call TXA a standard of care or to mandate that physicians give it. The absolute benefit is modest, if the subgroup turns out to be true, with a large confidence interval. However, with the new data (which is somewhat tricky because it is unpublished) I think it might be reasonable to use TXA in this subgroup while waiting for future research. My original conclusion still stands: read the trial and decide for yourself.
There was no change in the primary outcome. There was no change in mortality. There was no change in disability. Despite the incredible hype, there is nothing in this trial that indicates that TXA should be given to patients with an isolated head injury.
Personally, I think it is a mistake to change practice based on disease specific mortality, especially in a trial that was originally powered to show a difference in all cause mortality (although the trial did end up under-powered). There are certainly people who will use this trial to change their practice, and I think that is fine. Evidence based medicine is about taking all of the evidence available to you, and then using your clinical expertise to make the best decision for the patient in front of you. I will not criticize doctors for giving TXA to these patients, but clinicians should also not be criticized for withholding the medication. This trial clearly does not set a standard of care. This trial clearly should not be used to make strong guideline recommendations. Personally, I will not be using and will not recommend TXA for isolated head trauma until we see more research, but I suggest that everyone read the trial and decide for themselves.
October 31, 2019: Based on extra, unpublished data I have read, my conclusion from this study may change somewhat. That data is confidential, so I cannot say too much. However, in the main text of the post I do mention one thing that would change my conclusion. If I get permission to add extra numbers, I will update the posts as soon as possible.
There is another RCT of TXA in traumatic brain injury that randomized patients in the prehospital setting. Although the results have not been published anywhere that I can see, they are posted on clinicaltrials.gov, and it is also a negative trial with no benefit in terms of mortality or neurologic outcome. The fact that this negative trial was completed long before CRASH 3 but still hasn’t been published in a journal is an excellent example of publication or time lag bias.
I love the collaboration with FOAMed seen with this article. It was released early to a number of prominent members of the FOAMed community, so we were presented with many interpretations almost immediately. Over at EMNerd, Rory comes to the uncharacteristically mild conclusion that “we should not make the mistake of prioritizing its administration above other more important, more efficacious interventions.” On St Emlyn’s, Simon’s take is that TXA is “cheap and easy to give, it’s safe and effective”, and that we should be routinely giving it to most head injury patients (GCS 9 and above, or reactive pupils, or any extra-cranial injury). I think a key difference between my interpretation and Simon’s is because of prior-probability. He starts his post by saying TXA has been shown to save lives in postpartum hemorrhage and is well established in trauma, while I disagree with both those points. On PulmCrit, Josh is also taken in by the subgroups: “As a statistical rebel, I would consider this trial to be positive, despite having a technically negative primary endpoint.” On the SGEM, Ken is joined by Salim Rezaie, and their conclusion is “we cannot recommend the routine use of txa for patients with isolated traumatic brain injuries at this time.” Sal also has his own post on REBELEM, which concludes “the overall effect size of TXA on ICH is not that impressive based on this trial and also not statistically significant.” Looking at the subgroups, he says it would be wrong to conclude TXA is ineffective, “but it is difficult to make definitive conclusions when looking at these subpopulations.” The guys at the Resus Room podcast discuss the trial, including an interview with the lead author, and are very excited about TXA for these patients, and consider this a practice changing article, and hope that guidelines will change (which I think would be a mistake). On EM Literature of Note, Ryan acknowledges that we will probably see an increase in TXA use after this trial, but says “we certainly ought not be dogmatic about the necessity of its application in isolated head injury.” The crew at badEM also discussed the trial, with the interesting added perspective of applying this evidence in settings with different resources than many of us are used to, and their conclusions were mixed and nuanced but generally positive.
That is a pretty wide range of opinions from a single trial. Some of us are saying that we won’t give TXA, while others are suggesting that it should be given routinely. There are some very smart people on both sides of that spectrum. That might be frustrating to some people, but I think it is very telling. Science requires replication to gain clarity. In medicine, we love stopping after a single trial and either declaring that something works or that it doesn’t. Science doesn’t work that way. It doesn’t give us binary answers. We are never 100% certain.
The multitude of opinions clearly suggests uncertainty. It reminds us that we shouldn’t make definitive statements on the issue. That our guidelines should not be too strict, because we clearly don’t know the answer. That we shouldn’t criticize our colleagues if they happen to choose a different treatment option after reading this trial. For now, you just need to make the best decision you can, based on your personal medical expertise, the patient in front of you, and the numbers that have been presented here.
To keep up to date with all the latest critical appraisals, make sure to sign up for our email list:
Also, if you enjoy these posts, you can help us out by leaving us a 5 star review:
CRASH-2 trial contributors, Shakur H, Roberts I, et al. Effects of tranexamic acid on death, vascular occlusive events, and blood transfusion in trauma patients with significant haemorrhage (CRASH-2): a randomised, placebo-controlled trial. Lancet (London, England). 2010; 376(9734):23-32. [pubmed]
Roberts I, Belli A, Brenner A et al. Tranexamic acid for significant traumatic brain injury (The CRASH-3 trial): Statistical analysis plan for an international, randomised, double-blind, placebo-controlled trial [version 2; peer review: 3 approved, 1 approved with reservations]. Wellcome Open Res 2018, 3:86
Sprigg N, Flaherty K, Appleton JP, et al. Tranexamic acid for hyperacute primary IntraCerebral Haemorrhage (TICH-2): an international randomised, placebo-controlled, phase 3 superiority trial. Lancet (London, England). 2018; 391(10135):2107-2115. PMID: 29778325 [free full text]
The CRASH 3 Trial Collaborators. Effects of tranexamic acid on death, disability, vascular occlusive events and other morbidities in patients with acute traumatic brain injury (CRASH 3): a randomised, placebo-controlled trial. Lancet (London, England). 2019; PMID: 31623894 [free full text]
Wallach JD, Sullivan PG, Trepanowski JF, Sainani KL, Steyerberg EW, Ioannidis JP. Evaluation of Evidence of Statistical Support and Corroboration of Subgroup Claims in Randomized Clinical Trials. JAMA internal medicine. 2017; 177(4):554-560. [pubmed]
WOMAN trial collaborators . Effect of early tranexamic acid administration on mortality, hysterectomy, and other morbidities in women with post-partum haemorrhage (WOMAN): an international, randomised, double-blind, placebo-controlled trial. Lancet (London, England). 2017; PMID: 28456509 [free full text]
Morgenstern, J. CRASH 3: TXA is no wonder drug, First10EM, October 28, 2019. Available at:
13 thoughts on “CRASH 3: TXA is no wonder drug”
“…TXA might increase morality in some patients, which is a very important harm to consider.” Well said… 😉
Any thoughts on the MAtters (1 and 2) trial?
There will be a very long post coming out at some point reviewing all the TXA evidence for every indication. We are going to record an EMCases Journal Jam covering the topics. The problem with MATTERS is the nature of observational data. I will write more about it in the coming months.