Out of 22,349 stays, a tool integrated into the patient file equals case managers on the first day, but is twice as wrong as them 24 hours before discharge.
Published on September 3, 2026 in JAMA Network Open, the study compared, on 22,349 stays involving 17,173 patients at a Houston hospital, two ways of estimating the discharge date: that of a commercial AI tool integrated into the electronic patient record, and that of case managers in their usual work.
At the time of admission, the two methods are equal, and are equally mediocre: 4.20 days of average absolute error for the algorithm, 4.27 days for humans. Forty-eight hours before release, the AI shows 1.59 days of error compared to 1.29 for managers. Twenty-four hours before, the gap becomes clear: 1.93 days compared to 0.98. At this stage, managers fall within one day in 79.5% of cases, the tool in 37.9%.
The systematic bias tells the same story. The algorithm heavily underestimates the length of stay at entry (−2.74 days) and then overestimates as exit approaches (+1.24 days), while human estimates remain focused. However, it is precisely in the last forty-eight hours that forecasting is used to organize beds and transport.
Read the full article on the original source
Go to the source →