Brain death prediction based on ensembled artificial neural networks in neurosurgical intensive care unit

A novel ensembled artificial neural networks (EANN) model was used for brain death prediction in this study. The experimental material for this study focused on the severe head injury patients with different level of Glasgow Coma Scale (GCS) in neurosurgical and traumatic intensive care unit (ICU) of National Taiwan University Hospital (NTUH) in Taipei. Two prediction models were established, i.e., the first has 11 input variables while the second has 14 input variables. The root mean square error (RMSE) and the receiver operating characteristic (ROC) curve analysis were applied to examine the performance of the prediction models. Regression analysis was developed to compare the prediction results with EANN model. Besides, sensitivity analysis was carried out to evaluate the significance ranking of the input variables. For the 11 input variables models, the topologies of 11-5-1, 11-6-1, 11-20-1, 11-22-1, 11-11-10-1 and 11-20-10-1 gave similar results, which had the best performance with the value of area under curve (AUC) (0.96) and accuracy (0.9167), while the topology of 11-5-1 presented the smallest RMSE (0.2023 for test data). For the 14 input variables models, all of the topologies performed very well with all the parameters in ROC curve analysis were 1.0; however the 14-14-5-1 topology when choosing the best networks in ensemble model showed the best performance (RMSE for test data was 0.0423), which are much better than the results of regression analysis (RMSE for test data was 0.38 for 11 variables and 0.171 for 14 variables). With the help of sensitivity analysis, the significance ranking results for the 11 and 14 input variables in this study were presented. In conclusion, the successful results of RMSE and ROC curve analysis confirm that EANN model is useful to be applied to predict such a complicated issue as brain death.