Artificial neural network versus subjective scoring in predicting mortality in trauma patients

Current methods of trauma outcome prediction rely on clinical knowledge and experience. This makes the system a subjective score, because of intra-rater variability. This project aims to develop a neural network for predicting survival of trauma patients using standard, measured, physiological variables, and compare its predictive power with that obtained from current trauma scores.

The project uses 7688 patients admitted to the Swedish Medical Center, Colorado, U.S.A. between the years 2000-2003 inclusive. Neural Network software was used for data analysis to determine the best network design on which to base the model to be tested. The model is created using a minimum number of variables to produce an effective outcome predicting score. Initial variables were based on the current variables used in calculating the Revised Trauma Score, replacing the Glasgow Coma Scale (GCS) with a modified motor component of the GCS. Additional variables are added to the model until a suitable model is achieved.

The best model used Multi-Layer Perceptrons, with 8 input variables, 5 hidden neurons and 1 output. It was trained on 5881 cases and tested independently on 1807 cases. The model was able to accurately predict 91% patient mortality.

An ANN developed using pre-hospital physiological variables without using subjective scores resulted in good mortality prediction when applied to a test set. Its performance was too sensitive and requires refinement.