An Artificial Neural Network Derived Trauma Outcome Prediction Score as an Aid to Triage for Non-Clinicians

In mass casualty events Emergency Medical Service Providers (EMS) choose treatment at Scene or a "scoop and run" approach. The latter requires clinically trained personnel at the reception site to triage patients. Current methodology based on Revised Trauma Score (tRTS) requires use of Glasgow Coma Scale, a method reliant on experience and clinical knowledge. This makes the system subjective and often inadequate for non-clinicians. This project attempts to develop a simplified outcome prediction score using an artificial neural network for use by a non-clinically trained EMS to aid triage. The project uses National Trauma Data Bank, Version 6.1. Tiberius Data Mining Software created Neural Network models. Variables considered were values that could easily be obtained during an event. Binary values were used for low SBP and low Respiratory Rate, coded using the RTS scoring table as a basis, and age indicators. A modified motor component of Glasgow Coma Score was created to negate the need for clinical knowledge. Best performing models, identified by Gini coefficient and ability to predict mortality, were with 8 and 10 neurons. On mortality prediction all even numbers of hidden neurons have similar performances. Training sets were compared to test sets, and found to be identical in Gini coefficient and performance. Models performed well in predicting mortality compared to standard outcome predictors. Possible additional variables such as gender or ethnicity might improve the Neural Network predictive ability. Pulse appears an essential variable not recorded by the NTDB.