Should you get the blood transfusion?

Gur Huberman writes:

Apropos Ethics & Logistic Regression, the piece you wrote with Madigan:

In late 2001 I [Gur] broke my femur trying to rollerblade with my daughter. (No IQ award for that.)

I had surgery and my recovery was slow. Every time I tried to get on crutches I’d collapse and faint. Diagnosis: Anemia. Suggested response: Blood transfusion. Having lots of time in the hospital I read the warning associated with transfusion. As I recall it, it mentioned probability of 1 in 3 million of getting AIDS, 1 in 5 million of getting hepatitis, and 1 in 3,000 of having a hemolytic reaction which could range from allergy to death.

In my book, in this context, the only serious probability was the 1 in 3,000.

I refused blood transfusion. This lasted about a week until multiple doctors convinced me to accept the transfusion. I did, was given the transfusion in the evening and by morning I had no problem using the crutches.

Two days later I moved to a rehab hospital.

Within a few days I was diagnosed there with a life threatening pulmonary embolism. A possible cause was my prolonged immobility due to my weakness due to my refusal to accept blood transfusion.
What did I learn:

1. The warning text was probably written by lawyers, not doctors. Their objective was to protect the hospital, not the patient.

2. More clarity regrading the risks of omission might have made a big difference.

I like this story, partly because the way he tells the story, I was expecting the punch line to be that the transfusion was not necessary.

But the real message is that a 1 in 3,000 chance of dying is not so high! If you’re a 50-year-old man, you have a 1 in 200 chance of dying this year! (And the one-year time period seems like a reasonable comparison, in that this sort of issue doesn’t come up every day.)

30 thoughts on “Should you get the blood transfusion?

  1. i had malaria in the mid 80s in africa and took a drug (fansidar) that, it was told to me, had a 1 in 10,000 chance of killing me. i figured, ok, if i took that kind of risk every day (which i clearly didn’t) that would give me an expected 30 years. the drug cured me quickly.

  2. hate to be the actuary here, but the 0.5% chance of death is across the population. if you are a healthy 50M, your chance of death is much lower, an order of magnitude and then some. but point well taken, people do lots of things that generate mortality risks 1 out of X,000 where X is probably much smaller than they think (e.g. driving/riding in a car).

    • Here is something that confuses me: If someone asks “what is the probability I will die in a car accident?” Do you want probability of dying in a car accident, or probability of dying in a car accident given that you have not yet died in a car accident (hazard ratio)?

  3. I once heard a nurse say her mother was reluctant to take some new (to her) drug because of the warnings on the box. The nurse asked her, “Have you read the label on a bottle of aspirin [or was it acetaminophen?]?” The mother looked, thought a moment, and then took the pill.

    According to https://plus.maths.org/content/os/issue55/features/risk/index, that transfusion was only about twice as risky as giving birth in the USA.

    Given the variety of patients and illnesses, I imagine the risks of omission might be harder to estimate that the risks of the treatment.

  4. Today’s standards for transfusions are a lot more restrictive than in 2001. That is doctors are encouraged to transfuse less and allow a lower blood count than previously. The Cochrane group that generally does good work has done a meta analysis of multiple studies and did not find harm to patients from restricting transfusions; their analysis looked at thromboembolism specifically. There was actually some concern in the past that raising the blood count would raise the risk for thromboembolism based on observations of patients who have abnormally high hematocrits due to disease (polycythemia) and rheological mechanisms.
    Major trauma like a femoral fracture releases a large amount of procoagulant cytokines. Teleologically, your body responds to the injury by thinking that you are about to bleed to death.

  5. Removing, for a moment, my statistician’s hat but keeping my human hat:
    Is Gur Huberman OK??
    The story ended with him being diagnosed with a “life threatening…” which, to me, makes it more than just a good stats parable. (though it is that, too)

  6. This is just like all the cognitive illusion stuff. Sure, 1/3000 for the adult population is probably not a risk that a healthy person should take. You would be foolhardy to walk into a MD’s office and say hey give me a transfusion. But given that you he was not healthy and had other risks going on that the blood transfusion would potentially mitigate, worth the risk.
    This is why sometimes all the emphasis on how routine mammography (to name one example) is not beneficial never explains that the discussion is only for the adult population of women, not for the adult population of women whose mothers had breast cancer or who have other risk factors.

  7. It seems that making ethically sound decisions incorporating uncertainty is difficult not only because of the problem associated with “properly” quantifying uncertainty (as this example illustrates), but also because it is difficult to physically interpret uncertainty. For instance, consider the following variant of the trolley car problem: assuming that you can save an *average* of four people by switching the track that a train is currently on at the price of taking the life of one on *average*, what would you do?

  8. One thing I’m not so clear on though, aren’t there other methods of treating anemia? I mean, maybe doses of iron or whatnot, and, also, aren’t there other ways to get some exercise that would reduce your risk of being sedentary, like maybe using a wheelchair to move around, etc? I think the moral of the story here is that doctors don’t necessary look at the big picture and offer a range of alternatives. Doctors don’t do risk management very well in my humble opinion (on average).

    • Also, with regard to the “probable cause” I’d really like to hear some kind of quantitative estimate there. It’s easy to say “ooh if you’d only listened to us and gotten that transfusion you’d have been able to be active and not had that embolism” but the truth of the matter might be more like “you were destined for an embolism due to all the various clotting factors in your blood, and the transfusion probably even made that worse…” so I’d be careful of trying to learn too much from this story.

      • Transfusion is the only way of treating anemia immediately; giving iron takes weeks. There is no substitute for transfusion when immediate correction is required, but for many years until quite recently the criteria for giving a transfusion after a surgery when the patient is not in immediate danger were ad hoc, ie, rather by the seat of the pants. More recent actual scientific looks at the situation have resulted in fewer transfusions being given. I should say here that there is a clear need for transfusion in some circumstances, and I am a donor and invite others to also give.
        The risk of thrombosis after major bone fracture and surgery is close to 40% with around a 1% pulmonary embolism fatality rate. It is standard to give anticoagulant prophylaxis.
        For most humans the balance between bleeding and thrombosis is loaded toward thrombosis a situation which worsens with age.

        • Ok, good, so your point is that the embolism is not necessarily related to the lack of movement, and that the evidence actually suggests that giving transfusion isn’t necessarily that great an idea unless there is something very specific requiring it, hence fewer are being given now. So, the story:

          “I refused a transfusion out of fear from the risks, and I probably should have taken it because it would have helped me avoid a thrombosis/embolism due to inactivity” is as likely as not a fairy tale version of the real situation medically speaking.

        • Also “there’s no substitute for transfusion when immediate correction is required”. Neither of us know whether immediate correction was really required here, but the idea that because he got an embolism, immediate correction probably would have been the better thing to do here is…. looking pretty sketchy to me.

          typically with transfusion isn’t their an increased risk of clotting?

          http://www.urmc.rochester.edu/news/story/index.cfm?id=2293

          I mean given that kind of background, shouldn’t we probably infer here that if he *hadn’t* gotten the transfusion he might NOT have gotten the clot/embolism?

  9. Stasis is a risk; did not intend to imply otherwise. The Cochrane analysis of multiple studies showed no impact of transfusions on incidence of clot when more restrictive (contemporary) criteria are used versus the older triggers. The clot risks are in unprotected patients. I was aresident when the modern total hip surgery came along. A study was done using venograms which are invasive, often painful, but highly specific. Lo, there were a lot of clots. Part of the reason for underdiagnosis was that symptomatic disease was much less commom and often happens after hospital discharge as in this case. Clinical trials of prophylactic anticoagulant treatments are quite positive.
    One last comment. Actual harm from transfusion is quite uncommon, but due to the perversity of life the risk is much higher for unnecessary transfusion.

  10. I wish we had started creating massive repositories of raw data. e.g. Rather than some aggregate risk measure wouldn’t it be useful to know outcomes for (say) 50 to 60 year old anemics that had a transfusion?

    Does a doctor have any tools to estimate risks from the data on cohorts of matching patients?

    • This is a great observation, that the average risk is not that useful for an individual trying to make a decision for their particular case. Spiegelhalter and someone else I don’t know wrote this great book, The Norm Chronicles, where they make this point (“the average is an abstraction, the reality is variation”, may not be the exact quote), that knowing the average risk may not help you decide what your individual risk is. This is a particularly interesting question for me; in about seven years from now I will have to decide (probably very quickly, when the phone call comes) whether I want a kidney transplant or whether I will continue to live on dialysis. The doctors confidently tell me that a transplant is the better option, on average. But they can’t tell me what the best option is for me. The tools are available for calibrating my individual cost function, but reliable data are not, or at least the doctors seem to be unaware of it. This is an amazing failure in medicine (at least, in transplantation medicine), to understand the value of individual-level risk calculation. I get the feeling that part of it has to do with vested interests: a transplant surgeon always wants to operate, and hospitals are always looking to get more patients transplanted so they can do their clinical studies for immunosuppressive therapy.

      This person faced the same situation—the doctors (presumably) gave him a number, with no further calibration for his particular case.

      • Here is a follow-up question which interests me greatly: realistically, what can we expect from individual risk assessments? I agree fully that average risk assessments are not that meaningful compared with individual assessments. And, as available data continues to increase, we will see more individualized assessments (note the rapidly growing trend towards “personalized medicine”). But, most predictive models do not perform well on an individualized basis. Further, the estimated probabilities seem to be subject to a great deal of uncertainty. So, if you have a personalized risk assessment that says that for YOU the risk is 1 in 1,000 of something happening, that is only an average – an average of the possible risks that YOU face. How realistic is it that the 1 in 1,000 estimate will be sufficiently precise to be useful in an individual case?

        It seems clear to me that knowing that the risk of something is 1 in 500 in the general population and 1 in 1,000 for someone like YOU is a step in the right direction. But, isn’t there a tradeoff – isn’t the latter prediction likely to be more uncertain than the former (due to smaller relevant sample sizes as well as the inherent inaccuracy in any predictive model, namely that there is only an X% chance that the prediction for any individual will be correct?).

        Put differently, any model is likely to be more accurate on average, than for any individual case. But we know that individuals are not average, so when identifiable risk factors are known at an individual level, we can improve on the risk assessment compared with just using the average. But won’t this come at a cost of increased uncertainty? Managing that tradeoff seems to me to be a big issue.

        • Dale: Yes!

          In medicine your are extremely lucky if there is high quality data reducing the uncertainty on average risks – very lucky actually.

          I know folks want to know their risk just like they want to know the true value of unknown parameters – hey may I could trick a funding agency into giving me lots of money to do that, maybe call it “personalized medicine”? No actually need a better name because Stephen Seen is already debunking that … omnipotent risk quantification?

          On a serious note there are hopeful developments like FDA’s mini-Sentinel (google it).

        • Is there a middle ground possible? Don’t go down to the individual level, but to sub-groups. Doctors don’t say: transplantation is better if X, Y and Z; it’s just better, period. Also, there is no uncertainty accompanying that statement. Waiting time in Germany for kidney transplants, we are told, is 10 years (+/- what?). So currently they deliver an overly precise answer to the wrong question. Anyway, what do I know? I’m only talking about one specific case I know about; maybe in medicine there is more nuance. Probably the doctors themselves have a more nuanced understanding, but they can’t communicate it that easily to a patient.

  11. @Keith:

    Shouldn’t the studies already have the individual level data? It is just that publication comes with aggregation.

    The missing link seems a database to capture and expose the full datasets from multiple surveys.

    • @Rahul

      First even with individual level data, the randomization only makes groups (or strata if stratified) equal on average so it is very hard to get defensible estimates of sub-group treatment effects (and Stephen Senn has pointed out individual effects are not even identified in usual study designs.) Recall if it is a remarkably well designed and implemented study it may not have much more than 50% power to detect an overall effect – usual target is 80%).

      Now just because we can only reasonably learn about additive treatment effects with these studies is no justification to assume the treatment effect really is additive – but it seldom gets further than checking or exploratory analysis. You don’t always get what you want but these trails you might get something useful…

      > The missing link seems a database to capture and expose the full datasets from multiple surveys.
      That’s mini-Sentinel – but the mini refers to it being in a pilot stage – so don’t get sick anytime in the near future.

      If I recall correctly it has data from 99 million patients and hopefully in the not too distant future the European’s we send in another 99 million plus. http://www.mini-sentinel.org/default.aspx

  12. @Dale:

    Sure it may be easier to predict the average but what good is an average prediction to an individual?

    The only reason to use the average prediction is larger sample sizes?

  13. Rahul
    I don’t understand. Even “individual” predictions are averages – they are averages of the estimated probability of various outcomes. I agree that computing averages by subgroups is better than one overall average (assuming there is meaningful disaggregation of outcomes by subgroups), but the issue I am raising is that all estimates are averages and it doesn’t seem straightforward to me to tell which average is the “best” to use. In reality, don’t we have choices such as the following:
    1. the probability that treatment A will be successful is X% (an overall average)
    2. the probability that treatment A will be successful for a group of individuals that are somewhat like you is Y%
    3. the probability that treatment A will be successful for you is Z%?
    My question concerns the fact that all of these estimates are averages. The first averages over the whole population as well as over the possible values for X. The second averages over the subgroup as well as the estimates of Y. And, the third averages only over the estimated values for Z – but is likely to be more uncertain than the estimates of X or Y.

Leave a Reply

Your email address will not be published. Required fields are marked *