In-Hospital Mortality Prediction in Patients Receiving Mechanical Ventilation in Taiwan

Abstract

Few studies have used pooled data for more than 2 years and few have analyzed data for patients receiving mechanical ventilation in Taiwan. Objective To validate the use of an artificial neural network model for predicting in-hospital mortality in patients receiving mechanical ventilation in Taiwan and to compare the predictive accuracy of the artificial neural network model with that of a logistic regression model. Retrospective comparison of 1000 pairs of data sets processed by logistic regression and artificial neural network models based on initial clinical data for 213 945 patients receiving mechanical ventilation. For each pair of artificial neural network and logistic regression models, the area under the receiver operating characteristic curves, Hosmer-Lemeshow statistics, and accuracy rate were calculated and compared by using t tests. Global sensitivity analysis and sensitivity score approach were also used to assess the relative significance of input parameters in the system model and the relative importance of variables. Compared with the logistic regression model, the artificial neural network model had a better accuracy rate in 96.3% of cases, better Hosmer-Lemeshow statistics in 41.2% of cases, and a better area under the curve in 97.6% of cases. Hospital volume was the most influential (sensitive) variable affecting in-hospital mortality, followed by Charlson comorbidity index, length of stay, and hospital type. Compared with the conventional logistic regression model, the artificial neural network model was more accurate in predicting in-hospital mortality and had higher overall performance indices.