Using machine learning to predict near-term mortality in cirrhosis patients hospitalized at the University of Virginia health system

This study aimed to predict near-term mortality in patients hospitalized with cirrhosis at the University of Virginia Health System using two approaches — logistic regression and a long short-term memory neural network, with an emphasis on the latter. Using data from de-identified medical record entries for over 500 patient stays in the intensive care unit, a recurrent long short-term memory neural network was trained to predict whether a mortality would occur between 12 and 24 hours from the time of prediction. This was achieved by feeding the network a sequence of consecutive observations for each patient in the training set, thereby allowing the network to “look back” and identify increasing mortality risk based on changes in vital signs and laboratory results over time. Performance of the neural network model was compared to that of a logistic regression model and the Chronic Liver Failure Consortium Acute Decompensation Score, a benchmark scoring method for mortality risk in cirrhosis patients. Although the primary model continues to need refinement, results indicate that the neural network and logistic regression models both outperform the Chronic Liver Failure Score. As more data are made available, further improvement in the performance of the neural network is anticipated.