Prediction of fatality crashes with multilayer perceptron of crash record information system datasets

Despite the effort o f the authorities and researchers, there has been no sign o f decreasing in the num ber o f fatal crashes annually. To analyze the deadly collisions, researchers have focused on finding w hich factors affect injury severity, and thus m any crash prediction m odels for it had been developed. C om m only the injury severity is categorized into five different classes. Still, in m any studies, m inority classes like fatality and incapacitating injury w ere m erged so that the dataset becom es balanced, and the m odel can provide decent predictions. H ow ever, this approach does not help analyze the fatal crashes as they are joined w ith other types o f injury. Therefore, in this study, w e proposed a m ultilayer perceptron m odel for binary classification o f crash fatality. The m odel w as proved to be able to handle heavily im balanced datasets w hile providing decent perform ance. M oreover, a sensitivity analysis w as conducted on the input o f the m odel to estim ate the im portance o f crash-related factors.