A dynamic Bayesian network model for predicting organ failure associations without predefining outcomes

Abstract

Critical care medicine has been a field for Bayesian networks (BNs) application for investigating relationships among failing organs. Criticisms have been raised on using mortality as the only outcome to determine the treatment efficacy. We aimed to develop a dynamic BN model for detecting interrelationships among failing organs and their progression, not predefining outcomes and omitting hierarchization of organ interactions. We collected data from 850 critically ill patients from the national database used in many intensive care units. We considered as nodes the organ failure assessed by a score as recorded daily. We tested several possible DBNs and used the best bootstrapping results for calculating the strength of arcs and directions. The network structure was learned using a hill climbing method. The parameters of the local distributions were fitted with a maximum of the likelihood algorithm. The network that best satisfied the accuracy requirements included 15 nodes, corresponding to 5 variables measured at three times: ICU admission, second and seventh day of ICU stay. From our findings some organ associations had probabilities higher than 50% to arise at ICU admittance or in the following days persisting over time. Our study provided a network model predicting organ failure associations and their evolution over time. This approach has the potential advantage of detecting and comparing the effects of treatments on organ function.