Mar 15th 2007
From The Economist print edition
Computers may soon be better than kin at predicting the wishes of the dying
IN 1947 a psychologist called Theodore Sarbin made a controversial suggestion to a medical conference. He proposed that a doctor is really just a machine whose purpose is to make actuarial judgments about the best treatment for a patient. And not a very good machine, at that, for Sarbin also suggested that medicine would benefit if “we could replace [the doctor’s] eyes and brain with a Hollerith machine”.
It was a remarkably prescient vision. The idea that Hollerith machines (or computers, to give their modern name) might sometimes be better than doctors at deciding how to treat a patient is now universally accepted. A computer program is, for instance, sometimes used to recommend whether the horrors of chemotherapy are likely to outweigh its blessings.
When machines trespass into the area of medical ethics, though, hackles rise. Here it is not the doctor that is being second-guessed, but the patient’s relatives. The question is, if you were in a coma, whom would you more trust to come to the conclusion that you would want: your spouse or a machine?
David Wendler, of the National Institutes of Health in Bethesda, Maryland, and his colleagues have looked into this question. Their answer, just published in the Public Library of Science Medicine, is surprising. At the moment, both are equally reliable—but only the machines are likely to get better at it.
Dr Wendler’s study began last year, when his team reviewed all the experiments they could find that had attempted to test how well people predict the wishes of patients with life-threatening conditions. Some of these studies used real patients whose conditions might have led them to fall into a coma—when, obviously, they could not make the decision for themselves. Others employed surrogates who were asked to make “living wills” outlining their preferences for treatment (or the lack of it) in various hypothetical circumstances. The desires expressed by these patients, whether real or surrogate, were then compared with what those patients’ kin predicted the patients would want, and also with the predictions of unrelated people (doctors, for example) who might be called on to make the decision if kin could not be found.
Dr Wendler found 16 published reports containing almost 20,000 pairs of decisions. His analysis showed that kin and patient agreed only 68% of the time. When they did not agree, kin were more likely to recommend treatment when the patient wanted treatment withdrawn rather than mistakenly to recommend withdrawal. Surprisingly, the bias towards treatment was equally strong when the decision was made by an unrelated person such as a doctor.
Other research has suggested that the variable most reliably governing whether a patient would want the machine turned off is the “1% rule”. This is that people seem to want life-saving interventions if there is at least a 1% chance they will recover the ability to reason, remember and communicate. Less than 1%, and it is time to pull the plug.
Using that rule of thumb, Dr Wendler and his colleagues wrote a computer program that assesses the prognosis for a patient, based on the sort of clinical criteria that the studies had described to both patients and predictors. Only 12 of the 16 original studies contained sufficient detail to be used, but the result was remarkable. In these 12 studies, human predictors guessed the patient’s wishes rather more accurately than was true when all 16 were lumped together—getting them right 78.4% of the time. Dr Wendler’s program achieved an almost identical result—78.5%.
Since that result is based on a single criterion, the 1% rule, Dr Wendler reckons he can beat it by adding other factors to the program. Older patients may be less willing to accept heroic, invasive surgery than younger ones; men might think differently from women; professors imagining themselves with advanced dementia may more readily turn down pneumonia treatments than dancers would. Dr Wendler’s guess is that by studying such preferences in more detail and adding them to the program, he might increase its accuracy by as much as another ten percentage points.
At the moment, such data do not exist. No one has yet had a reason to collect them. But they do have a reason now. The decision about when to pull the plug on a patient who is not expected to recover is unlikely ever to be handed over completely to a machine. But when no kin can be found, the program’s opinion might help. And even when a dying patient is surrounded by people who care about him, those people may welcome some guidance about what his wishes were likely to have been. Individuals are, indeed, individual. But that does not mean their dying wishes are all that different.