Illness severity scores have demonstrated poor generalization performance in forecasting mortality, particularly for patients at the highest risk of mortality. One of the drivers of this discrepancy is the preponderance of patients at low risk of death in training cohorts and the inflexibility of predicting mortality risk across the spectrum of disease severity with a single, inflexible, linear model. We explored a two stage, sequential modeling approach, and compared it to using a simple linear model as well as a state-of-the-art machine learning method.
Contained herein are all of the original code files used to produce this work. Please contact me at cvc@mit.edu with any questions.