Comparing ANNs to logistic regression for predicting trauma survival

A comparison of the abilities of two methods, TRISS (trauma injury severity score, a logistic regression method) and backpropagation ANNs, to compute survival probabilities for trauma victims has been presented. The intention is to find reasons for unexpected outcomes and so to improve treatment techniques. Using the same inputs as the TRISS model, the ANN exhibited superior performance in predicting deaths, mathematical expectation of the number of survivors, and overall prediction of life/death response. The results also show that there is still a (perhaps unacceptably large) number of unexpected deaths, by both models. Other work currently underway suggests that these results may be improved by refining the current inputs or including new inputs. For example, it appears that using actual age rather than the categorical age used in the models discussed here may provide significant improvement. Other analyses indicate that the patient´s sex and the involvement of drugs or alcohol affect outcomes and should be included in the model. Future work in this area will be concentrated on improving the ability of the ANN to predict survival and to investigate other technologies for improving results. In particular, other ANN architectures may be tried and the data mining technique of rough sets may be applied to the data set described here