Recently, I was doing a project on cancer hospital special quality metrics - a program called "PCHQR." These "cancer quality metrics" include "proportion of patients who die of cancer admitted to ICU" and "proportion who receive chemotherapy in last 14 days of life." These remind you of the more general topic of healthcare intensity and resource use, for end-of-life patients.
This week, NEJM has two articles about machine learning and big data in medicine - here, interviewing Obermeyer, and here, interviewing Hightower and Kohane. In the first interview, Obermeyer gets a chance to discuss his important 2018 paper in Science on end of life spending.
The lesson of Einav et al. 2018, (including Obermeyer) is easy to remember, but I don't recall hearing it before.
Yes, one-quarter of Medicare spending is in the last 12 months of life, and that number is easy to measure.
However, using big data to establish mortality predictions (do you know if the patient is in the last 12 months? or later?), they found that only 5% of Medicare spending is on patients who at the time have a greater than 50% mortality prediction.
The authors concluce, this goes against the cliche' that "a large share of healthcare dollars is wasted on small gains on those certain to die."
