Using neural networks for trauma outcome evaluation

The development of effective methods for predicting the survival or mortality of patients is a major focus of trauma research. We address this problem by applying neural network modelling, using as inputs established descriptors of physiologic and anatomic injury severity. We use two different "backpropagation" techniques with a variety of training strategies and examine the importance of training and test set composition, performance evaluation criteria, and interpretation of network outputs. The results of our study indicate that neural networks show significant promise for trauma patient outcome evaluation.