Using Machine Learning to Predict Mortality: Demystifying U.S. Healthcare Spending

Approximately one in every five dollars spent by Medicare is spent during a patient’s last twelve months of life, with most of this spending used to pay for inpatient medical care and physician costs. Despite taking up 20 percent of its budget, end-of-life cases account for just five percent of total Medicare beneficiaries. Should we conclude that the government is spending too much on hopeless cases where death is inevitable? New research suggests that the answer is no: Those who will die, prior to death, are indistinguishable from those who might die. Medicare spending tends to be high among high mortality conditions. Very sick patients require more care. While some patients die after significant spending, many do not, and a full understanding of healthcare spending in the U.S. should account for this fact. Health spending can be complex; so complex, that researchers have begun to turn to machines to understand it better.

In a study recently published in Science, researchers from Harvard, MIT, Stanford, and the University of Chicago showed that the large amounts spent by the government on the last few months of life are not a sign of inefficiency. The researchers followed a two-step methodology. First, they used a machine learning approach to predict the probability of dying among a random sample of Medicare enrollees given various information about the individual in the sample. Next, they analyzed how much of total Medicare spending was dedicated to those patients with the highest mortality probabilities.

In general terms, predictive machine learning models are algorithms that can be used to get a computer to analyze information, identify patterns among variables, and decide which specific patterns can be used to make the most accurate assessments. In this case, the researchers sampled 20 percent of all Medicare enrollees (about six million people) in January 2008. For each patient, they obtained Medicare claims, including data on inpatient care, outpatient care, and physical services, as well as all recorded health diagnoses. Then, they used a combination of three machine learning techniques commonly used to predict mortality on one-third of their sample to construct a predictive model. In their model, the variables with the highest predictive power were those that indicated the presence of a chronic condition and those that measured medical utilization—such as number of inpatient visits and ER visits. Finally, the researcher applied the model to the entire sample and obtained, for each person, an estimated probability of dying within the following 12 months.

It turns out that death is very unpredictable. As the study put it, “There is no sizable mass of people for whom death is certain (or even near certain) within one year.” Less than ten percent of those who died in 2008 had an annual mortality probability higher than 50 percent. In other words, for those who the model predicted a better chance of dying than living through the year, only one in ten of those people actually died. Even for the subset of those that were admitted to the hospital, only four percent of the patients that died had a mortality probability of more than 80 percent. The researchers estimated that the riskiest one percent of the patients accounted for less than $19.3 billion of the $386 billion spent in total by Medicare in 2008. Their estimations revealed that the riskiest one percent of patients—those with a mortality probability of over 46 percent—accounted for less than five percent of total spending. The researchers concluded that there was no evidence of a disproportionate allocation of resources to the riskiest patients. What they observed instead was that expenditure tends to increase with predicted mortality, which is only a reflection of the fact that health care costs increase with severity of sickness.

The researchers determined that although there is very high Medicare spending on the last twelve months of life, there are only a few individuals for whom death is near certain. The study noted that these results should not be understood as evidence of the absence of inefficiencies in Medicare. The paper suggested that further research should focus on the quality of care for very sick patients, as well as on the impact of specific treatments on survival rates and treatment of symptoms.

From a policy perspective, this study exemplifies how machine learning can be used to assess highly complicated social programs associated with large amounts of data. Machine learning played a key role in this study by analyzing millions of possible patterns among hundreds of health variables and deciding which ones produced the best model. This result would have been impossible to achieve without learning algorithms and demonstrates the high value of this approach for policymakers.

Article source: Einav, Liran, Amy Frinkelstein, Sendhil Mullainathan, and Ziad Obermeyer. “Predictive modeling of US health care spending in late life.” Science Vol. 360, Issue 6396 (2018): 1462-1465.

Featured photo: cc/(KatarzynaBialasiewicz, photo ID: 939771318, from iStock by Getty Images)