Survivorship bias in the ICU

A 99-year-old woman is admitted to the ICU. You shrug your shoulders, assuming such an elderly person cannot possibly survive acute illness. However, you’re surprised when three days later, she has stabilized and is successfully transferred to the floor, awaiting discharge.

The modern world, and certainly modern medicine, is full of opportunities to be wrong.

Many factors contribute to wrongness, but the most pervasive arise from our own flawed processes of reasoning. These are cognitive biases, systematic errors in how we reach conclusions that occur again and again, like bugs in our mental software. They cannot easily be eradicated, but with proper preparation and forewarning, we can at least be aware of their existence.

In the ICU, one common error is survivorship bias.

What is it?

Survivorship bias is the following “bug”: in any situation where a force is acting to remove individuals from a pool of people, events, or phenomena, the only population that remains will be those that have already survived the aforementioned winnowing process. Thus when we observe them, we are observing a non-random sample.

This seems obvious at first, but the consequences are not. The link above describes a classic example: upon examining World War II bombers who had survived attacks from ground-based gunnery, military planners observed a preponderance of damage to the wings, body, and tail of the aircraft. The natural conclusion is that those areas generally accumulated the worst damage, and hence should be armored more heavily. Right?

Wrong. If the entire pool of bombers could actually be assessed, that reasoning might be reliable. But the bombers that were destroyed by their damage didn’t make it home to be examined. The ones that did, the survivors, were actually demonstrating where a bomber could accrue damage without being destroyed. The areas to add armor were those without bulletholes.

A small twist, but one that reverses the conclusion. How can we apply this principle in the ICU? Let’s consider three examples.

The case of the striking anemia

Every provider has admitted a patient with GI bleeding. One important datapoint in such presentations is their degree of anemia on initial labs. Following the principle that lower is worse, and that patients without enough blood will die, a hemoglobin of 11 g/dL (mildly depressed) seems much less concerning to us than a hemoglobin of 6 (significantly reduced), and a hemoglobin of 3 seems to denote a true emergency.

But wait. Consider the survivorship bias.

We worry about a hemoglobin of 3 because we assume it portends a large, brisk bleed, and that without aggressive management it will soon be 0. However, a little common sense should be applied. A large, brisk bleed may cause a hemoglobin of 2 g/dL, but only for a few minutes; like speeding through a tollbooth, it will be whooshing past that result en route to rapid exsanguination. If the patient has voluminous hematemesis, severe hypotension, and is actively being intubated and massively transfused, that may be the correct context. However, in a moderately-symptomatic but generally stable person, severe anemia probably tells the opposite story. They cannot be bleeding rapidly, or they would not have made it to the hospital (or at least would look very obviously moribund). The very low hemoglobin actually suggests the bleeding must be slow, a subacute or chronic process, allowing the patient time to compensate. A patient who appears to be in shock with a hemoglobin of 10 might be far more ill.

(This should be taken with a grain of salt, of course; bleeding can be acute on chronic, and relatively rapid bleeding can occur in the setting of a longstanding anemia. It is also true that such patients will have less reserve, because you do need some red cells. Still, the point stands.)

The case of the extremely elderly patient

When a very old patient is admitted to the ICU, we note that age is a risk factor for poor outcomes from virtually every disease. However, reaching an advanced age also says something about the characteristics of the patient, which we may forget.

It is a truism that upon reaching middle age or thereabouts, most patients stop behaving like their calendar age, medically speaking, and start to instead resemble their state of health. There are 50-year-olds who look 100, and 90-year-olds who look 70. We understand this, and when we admit a 60-year-old man with chronic diseases of multiple organs, we acknowledge they have nothing in common with a person of the same age who still runs marathons. Indeed, most of those comorbid individuals do not survive to extreme old age; there are not many 90-year-olds with uncontrolled diabetes, CHF, renal failure, and so on.

Consider the consequences of this. In many ways, very elderly patients (over 85, say—certainly in their 90s and above) are the healthiest patients we see! This is classic survivorship bias. These patients aren’t immortal, and aren’t immune from the problems suffered by others. But if they develop a new medical problem, they succumb to it; they don’t endure it. Few serious diagnoses are “chronic” in a 100-year-old, because they die from them. The survivors are those who, by virtue of luck or lifestyle or genetics, do not often get sick.

It can be easy to invert this lesson, and assume that such folks are made of “tough stuff.” They survived two world wars, the Great Depression, and polio, so by gosh, they can probably survive pneumonia. This isn’t quite right. They’re not healthy because they’re old; they are old because they’re healthy. The more correct conclusion is that they should be viewed as being younger than their calendar age. A 90-year-old person in good health may be more like an average 60-year-old.

For some reason, I find that we understand this equivalence at younger ages, but once someone hits 90, we only see the number.

The case of the large pericardial effusion

By now, you should be getting the idea.

You place an ultrasound probe on a patient’s chest, and notice a very large pericardial effusion. “Is this effusion large enough to cause cardiac tamponade?” you ask, and in your head, the size argues in favor of “yes.”

However, the opposite may be true. Effusion size is correlated with tamponade physiology in only the most rudimentary way. Chronic effusions that develop over months or years give the pericardium time to expand, making very large collections possible without any compression of the heart. Conversely, an effusion of only 100 ml that accumulates over hours (from hemorrhage, for instance) may cause immediate tamponade. In other words—with all the caveats noted, such as the possibility of an acute-on-chronic situation—a large effusion may be an argument against tamponade, not in favor of it.

Conclusions

It’s easy to be wrong in medicine, but we can debulk the risk of it by recognizing the common threads in our wrongness. By understanding the true impacts of chronicity, self-selection, and survivorship bias in the diseases we see, we’re less likely to assume foolish things. And that’s always a nice goal.

Leave a Reply

Your email address will not be published.