It’s difficult to make long-term predictions about which patients will die, but some worry that more accurate tools might affect healthcare spending.
If you want to talk about wasteful spending at the end of life, health policy expert Bruce Jennings, M.A., suggests looking back to the 1960s.
Back then, much of the medical technology we take for granted today was brand new — and much of it was overused.
Nowadays, physicians have learned that many technologies, such as artificial life support, are often used inappropriately.
“People don’t have any idea how burdensome it is to have continued life support for a patient who is actually trying to die, and the technology is not allowing them to die,” Jennings, an adjunct professor at Vanderbilt University and a senior fellow at the Center for Humans and Nature, told Healthcare Analytics News™.
Avoiding extensive life-sustaining interventions in cases with little or no chance of survival isn’t only good for patients and families; it’s also good for the healthcare system.
“Right now, we save a lot of money by giving people the exit option,” he said. “But it’s probably not enough because too many ordinary people and families have the rights, on paper, to refuse treatment, but they can’t bring themselves to exercise those rights.”
The idea that people are opting for unnecessary or unlikely-to-succeed treatments when the end is near adds up to tens of millions — perhaps billions — of dollars in healthcare spending each year. Indeed, spending in the last year of life has become a hot-button issue in healthcare, premised on the idea that much of the money we spend at the end of life is simply ‘wasted.’
The data appear to tell a compelling story. In 2014, 2.6 million people died in the U.S., and four of every five of those people were on Medicare, according to a 2015 report from the U.S. Centers for Disease Control and Prevention. Medicare spent $600 billion on patient care that year, and approximately one quarter of that total was believed to have been spent on patients in the final year of life.
Such data have led to this idea: If the country wants to get its healthcare spending under control, it ought to take a look at curbing end-of-life spending.
But scientists at the Massachusetts Institute of Technology, Harvard University and Stanford University recently took a different approach. They sought to use a machine-learning model to analyze Medicare and mortality data to find out how many patient deaths were predictable and how much was spent on patients whose deaths seemed to be written in the statistics.
The results were conclusive: It’s really, really hard to predict death. Even in cases where a patient had a higher likelihood of death than survival, the corresponding spending was relatively small. In total, less than 5 percent of spending went patients whose predicted mortality was above 50 percent, according to the paper.
In other words, healthcare spending helps people who are sick. People who are sick sometimes die, and therefore people who die account for a significant amount of healthcare spending. But that doesn’t mean spending on the sick — even the very sick — can be counted as waste, the authors argue.
“Our results suggest that spending on the ex post dead does not necessarily mean that we spend on the ex ante ‘hopeless,’” the study concludes.
Jennings sees the findings as an important counter-narrative to the idea that end-of-life spending is fraught with waste. Rather, the report suggests that a high percentage of these people have a legitimate shot at survival, at least statistically.
“If we change the major narrative to… a lot of them have a chance at recovery, it doesn’t seem like too much to spend,” he said.
Although the study answers questions about our end-of-life spending, it also raises questions about the role predictive modeling could play in the healthcare decisions of tomorrow.
Amy Finkelstein, Ph.D., an MIT economist who co-authored the study, said she doesn’t expect computer modeling to ever accurately predict death in a way that would allow for the reigning in of end-of-life spending. For one thing, the study indicated that even with the availability of richer data, such as electronic health records, the predictive model fell short at deciphering which patients wouldn’t survive.
But Finkelstein said even highly accurate modeling wouldn’t easily translate into easy cost savings.
“We do an exercise in the paper where we imagine a hypothetical, better predictor. We produce an artificial ‘oracle’ predictor that puts some weight on our prediction and some weight on realized outcomes (i.e. increasing predicted probabilities if ex post you died),” she said. “Doing so, we can jack up the performance to well above anything in the literature and still not make a dent in the results.”
One reason why end-of-life spending is difficult to curb is that it’s not primarily the result of last-minute interventions. A well-publicized 2016 study in Health Affairs found that most patients don’t see a spike in spending at the end of life. In fact, nearly half of the Medicare recipients in the study had persistent high spending, generally due to having multiple chronic conditions rather than a specific disease. Another 29 percent had “moderate persistent spending,” meaning there was no significant increase in the final year of life, but their spending was less than those with multiple chronic conditions.
Only 22.3 percent of the patients in the study had a significant jump in healthcare spending during their final year in life. Of those, 12.1 percent saw a last-minute increase, as opposed to a steady rise.
“These findings suggest that spending at the end of life is a marker of general spending patterns often set in motion long before death,” concluded the authors, led by Matthew A. Davis, Ph.D., M.P.H., of the University of Michigan.
Predictive modeling is getting better and better. While it might not be able to predict long-term deaths with sufficient accuracy, there are signs that its short-term capabilities are impressive.
A paper published earlier this summer in NPJ Digital Medicine looked at the possibilities of deep-learning models based on data from electronic health records. Among other findings, the study suggested that the deep-learning model outperformed the augmented Early Warning Score (aEWS) at predicting mortality at 24 hours after hospital admission.
“If a clinical team had to investigate patients predicted to be at high risk of dying, the rate of false alerts at each point in time was roughly halved by our model,” wrote the authors, who represented the University of California San Francisco, the University of Chicago, Stanford University and tech giant Google.
In one case, the aEWS model gave a cancer patient with fluid in her lungs a 9.3 percent chance of dying while in the hospital, but the deep-learning model rated her chance of death during the hospitalization at 19.9 percent. The patient died 10 days into her stay.
Finkelstein, at MIT, said if prediction models were able to become much more accurate, it would require society to evaluate how it wants to deal with people with low chances of survival. That’s more of a philosophical and ethical question than an economics one, she said.
For Jennings, this brave new world could prompt questions like this: What percentage of survival is too low to be worthy of healthcare spending? The answer could vary from society to society and might even depend on the relative wealth of a country, he said.
There are risks in trying to set a standardized threshold of which patients receive care and which don’t.
If, for instance, a policy was enacted that people with less than a 37 percent chance of survival were not entitled to expensive end-of-life interventions, it would be imperative that physicians agreed with the threshold, Jennings said.
“How you can get the practitioners of medicine to obey the rules that the policymakers lay down?” he said. “If you make those rules too stringent, you violate the conscience of physicians.”
Any policy that encouraged physicians to violate their consciences would be bad for medicine. If physicians reacted by trying to game the system, that would be bad too.
Generally speaking, end-of-life decisions have been left to patients and families in the U.S., and they sometimes refuse life-sustaining treatment because it doesn’t accord with their values or wishes. That kind of model is good for patients and the economy, Jennings said.
However, the use of machine-learning and artificial intelligence technologies as guides for end-of-life decisions ought to be carefully thought through, he added.
“It is a different thing to use artificial intelligence capabilities to make determinations and set rules about what individual physicians and insurance companies will facilitate going forward,” he said. “I think computers can help us very much understand data. I don’t think that computers ought to set the rules for extremely complicated life decisions.”
Finkelstein said for now the lesson from her study is that if “wasteful spending” at the end of life is an issue, it certainly isn’t an economic emergency. Rather than trying to decide which patients are worthy of spending, she said, we ought to try and learn which types of treatments are worthwhile.
“We should stop focusing on calling spending that occurs at what is ex-post the end of life ‘wasteful’ and do the hard work of identifying interventions and types of care that actually seem to produce no benefits to patients,” she said.
Get the best insights in healthcare analytics directly to your inbox.