Ensembled neural networks for brain death prediction for patients with severe head injury

The concept of organ donation has gradually been accepted by people in recent years so the judicial brain death determination process becomes very important. Clinically, patients with irreversible apnoeic coma (IAC) will be considered legally as brain death based on a judicial process, but this process can only be applied to people who had already signed the letter of consent to organ donation. The main idea behind the proposed model is to find out an easier way to diagnose the prognosis of patients with severe head injury, and offer the medical staffs more information to determine brain death. Therefore, the technique of ensembled neural networks (ENN) based on multi-layer perceptron (MLP) network has been applied to construct the prediction model of brain death index (BDI). Ten different signals were chosen to be the input data. Using these ten parameters, medical doctors depend on their experience to score the BDI hourly values. The BDI values from medical doctors become the training target of the ANN training process and the standard index of testing process. Moreover, in order to compare the differences between doctors’ and the network's rankings for the input data, the ranking of order of precedence of each input signal is analyzed via sensitivity analysis. The results show that the 4 layers network with validation has better performance than 3 layers. For sensitivity analysis, most of the input variables’ ranking from trained model were similar to the ranking of the medical doctors except RR/RR(Set) this parameter and 4 other parameters (PS-R, PR-R, PS-L, and PR-L) are difficult to rank, even medical doctors cannot decide the ranking accurately. Using the best topology structure of MLP 10-10-5-1, the ensemble neural network could effectively predict the BDI with small errors (i.e. training error = 0.219087; validation error = 0.370485; testing error = 0.280515). In conclusion, this model can provide medical staffs a reference index to evaluate the status of IAC and brain death patients. However, more clinical data are still needed, perhaps to refine the weights of EANN, and certainly to see how widely the model is applicable.