Noticing that a patient on the ward is going downhill · Contested (since 2021)

In June 2021 an external validation in JAMA Internal Medicine found that the sepsis model built into the Epic record, running in hundreds of American hospitals, identified only 33 percent of sepsis cases, with an area under the curve of 0.63.

The model had been sold and deployed without independent validation, and when one was finally done on 27 697 patients it caught a third of the cases while flooding clinicians with warnings. The opposite result exists. Kaiser Permanente's Advance Alert Monitor, reported in the New England Journal of Medicine in November 2020 across 21 hospitals, was associated with lower mortality, and the difference was not the mathematics but the design: it summoned a trained response team instead of raising a flag in a chart. Reading the numbers that precede a deterioration is now a machine task. Deciding that a patient is not right, walking into the room and doing something is not, and the Epic case shows what an alert is worth when nobody has been made responsible for answering it.