This study applies artificial neural networks to examine North Carolina motor vehicle crash data with the aim of identifying the factors influencing crash injury severity in alcohol-related crashes. A 3-Class neural network model and a binary (injury and non-injury) neural network model were developed using 49 variables with 63.58% and 68.22% validation accuracies respectively. Overturning was found to be the topmost factor influencing fatal and severe injuries in both the models while wearing a lap belt and deployment of airbags were found to be important factors contributing to a non-injury. Interestingly, the results suggest that driving a passenger car while intoxicated possibly leads to less severe injuries than driving a SUV under the same conditions. Driving in a residential area was found to be an important factor in a non-injury crash compared to driving in a commercial area. The accuracy of the 3-Class and binary models was improved using multiple neurons, and the optimum validation accuracies of 65.33% and 69.65% were obtained with the addition of six neurons in both the models.
Banerjee, S., & Khadem, N. K. (2019). Factors Influencing Injury Severity in Alcohol-Related Crashes: A Neural Network Approach Using HSIS Crash Data. ITE Journal, 89(3), 42–49.