Application of Multiple Neural Networks to Time Sequence Data - Prediction of Nosocomial Infection in Intensive Care Unit Patients

We have invented the method of modelling time sequence data for prediction using multiple neural networks (NNs). In this study, we examined whether multiple NNs outperforms logistic regression and single NN in the prediction of nosocomial infection in intensive care unit patients (n=16,584). The three predictive models were developed using the 80% training subset and their predictive performances were assessed using the 20% testing subset in terms of classification accuracy and the area under a receiver operating characteristics curve. Overall the highest predictive performance was found in multiple NNs, followed by logistic regression and single NN. The predictive performance of multiple NNs was kept at a constant level, whereas that of logistic regression and single NN decreased with increasing a follow-up period.