Investigating Injury Occurrence in Motor Vehicle Collision Using Artificial Neural Networks

Abstract

Vehicle collisions amount to a significant loss of life in America. Upward of 30,000 lives are lost each year in the United States alone. Hundreds of millions of dollars are spent each year on vehicle safety and roadway design in the United States, and many studies have been conducted on the causal factors for vehicle collisions. This study used artificial neural networks as a means to predict the occurrence of injury of a vehicle collision. The objective was to classify the levels of the “injury occurrence” in vehicle collisions, given certain input variables, including demographic and environmental factors involved in crashes. Using Neural Ware's Predict software, a back‐propagation neural network model was constructed, trained, tested, and validated using data from the 2006 and 2007 Florida Traffic Crash Database. The neural network's computational power was iteratively increased by adding hidden layers to boost its predicting power until the performance began to plateau. Classification rates as well as Kolmogorov‐Smirnov (K‐S) statistical analysis were used to validate the model. A sensitivity analysis was then applied to find the level of contribution that the input variables had on the injury occurrence. Accident severity, number of lanes, and road surface conditions were found to have the most positive impacts on the injury occurrence. The implications of the study are discussed at the end of the article.

Investigating Key Factors Influencing the Severity of Drivers Injuries in Car Crashes Using Supervised Machine Learning Techniques

“…However, the area under curves for ROC plots shows very close prediction accuracy for class 1 and 3 (no injury and fatality), and less accuracy for class 2 (injury), this could be due to the fact that minor and major injuries were used undistinguishably with small dataset size for major injury cases. However, the overall prediction accuracy is considered acceptable and promising compared with previous literature such as [11,22,28,45]. In these studies, the ANN prediction accuracy ranged between 55% and 65%.…”

“…Several studies used this particular type of ANN to investigate crashes risk factors. To name a few, [22] used a feedforward backpropagation ANN model to predict the occurrence of injuries in vehicle collision crashes in Florida. It was found that number of lanes and road surface conditions contribute the most to the injury occurrence.…”