查看文章

Early identification of patients at risk for deterioration and death can allow clinicians to appropriately allocate resources and provide appropriate care to those most in need. We have developed a machine learning algorithm for 48 hour all-cause mortality prediction. Algorithm performance has been validated in a retrospective setting, and its effect on patient outcomes is being evaluated in a prospective randomized clinical trial. We used an ensemble of boosted decision trees to create the prediction algorithm. For retrospective validation, we trained the algorithm on 94,222 encounters from the University of California San Francisco (UCSF). We also performed cross-population validation experiments on 13,668 encounters from University of Washington Medical Center (UW), and 77,142 encounters from Stanford Hospital. The algorithm was trained and tested for mortality prediction at 12, 24, and 48 …