aybe you have heard the refrain before: The U.S. spends too much money on the dying. Every year, 5 percent of Medicare beneficiaries die, but one-quarter of spending occurs in the last year of life. Side by side, these stats have fed a widely held belief that, in an exorbitantly expensive health care system, much of end-of-life care goes to waste.
A new study, published in the journal Science, pushes back on this notion. The researchers, a team of three economists and one physician, used machine learning to predict mortality and re-examine spending. In their new estimate, patients with the highest one-year mortality risk account for less than 5 percent of spending, much less than the original one-quarter claim.
But the conclusion that most surprised author and Massachusetts Institute of Technology economist Amy Finkelstein: Death is highly unpredictable.
Before their analysis, Finkelstein and her co-authors at Stanford, Harvard, and the National Bureau of Economic Research thought that patients who die within the year would have had extremely high mortality risks at the time of hospital admission. Instead, the data revealed that patients in the group with the steepest risk still were slightly more likely to survive the year than not. Even with rich data and sophisticated algorithms, predicting life and death has odds similar to flipping heads or tails.
“We spend money on sick people — some of them die, some of them recover,” Finkelstein said in an interview. “Maybe some recover, in part, because of what we spent on them.”
If neither physicians in the hospital nor economists behind the scenes can predict death with some degree of certainty, are resources really wasted on providing care?
Implicit in the dialogue about wasting money in the last year of life is the assumption that there is a good way to distinguish the sick and dying patient from the sick patient who will survive.
Imagine you paid $5 at a parking meter on a Sunday, unaware that street parking is free on weekends. If you happened to miss the giant, neon sign that reads “free parking,” then you wasted five bucks. If, on the other hand, there was no such sign, then it would have been impossible to know, when feeding the meter, that your payment went to waste. Similarly, if a physician who opts to keep a patient in the ICU on a ventilator for $1,500 per day is unable to predict her patient’s fate, the care is less decidedly inefficient.
“Just because someone is seriously ill with an uncertain prognosis doesn’t mean that their health care spending is wasteful,” said Dr. Stephanie Harman, who works in palliative care and biomedical ethics at Stanford.
Harman was not involved in the new research but has also applied machine learning techniques in her field to identify patients who may benefit from palliative care interventions. This new study, she said, underscores the key problem that arises when working backward from known deaths — researchers erroneously assign terms like “wasteful” due to the inherent bias in a retrospective approach. That $5 only seems like a waste once you realize that parking was free all along.
As a general rule, the sickest patients require the most expensive interventions, and a large fraction of the concentrated spending before death can be explained by how seriously ill these patients are. The study’s authors estimate that this fact accounts for 30 to 50 percent of spending in the last year of life.
The authors acknowledge possible shortcomings in their research. They wondered whether their algorithm — generated using data from millions of Medicare enrollees — fell short. “Maybe we just suck at predicting,” said Finkelstein.
To address this concern, the team test-drove a more accurate model. They built an “oracle,” a machine learning algorithm that weighs both real deaths and predicted deaths. Even under this superior model, the data barely budged — patients with high mortality risk still accounted for only a small fraction of spending.
One implication of these findings is that evaluations of end-of-life care need to look at more than just spending and whether a patient dies. “We need to also consider the quality of the care that’s delivered and the patients’ quality of life,” Harman said. Finkelstein said it is time to begin the challenging but critical work of combing through, “intervention by intervention,” to determine which policies, procedures, and treatments produce health benefits and which, unfortunately, do not.
There is undoubtedly tremendous waste in the medical system. Dr. Atul Gawande drew attention to the problem in a 2015 New Yorker article, which examined the forces that contribute to a medical culture of “overtesting, overdiagnosis, and overtreatment.” The waste is real, but Finkelstein said maybe cutting care in the last year of life is not a fruitful way to clean it up.
Quoting the movie “The Princess Bride,” with a mixture of humor and sincerity in her voice, Finkelstein said, “‘I do not think it means what you think it means.’ We need to be a little more careful and a little more sensible when leaping from fact to conclusion.”