Prediction of Traffic Crash Severity – A Solution by International Data and Deep Learning Models

Road traffic crash (RTC) is one of the critical problems worldwide. According to WHO, millions of people are killed and injured in RTCs. The global mean fatalities due to RTCs was 18.2 deaths per hundred-thousand people in 2016, with lower-income countries suffering the most. About 1.35 million people face death in traffic crashes annually meaning a daily average exceeding 3,290 people. Casualties involving injury of impairment coming to a sum between 20 to 50 million. Traffic crashes come in the 8th place as a death leading cause, summing up to 2.2% of death counts worldwide. Countries of low and middle income owning below 50% of the vehicles of the world contribute with more than 90% of the crash fatalities. The accurate prediction of traffic crash severity contributes to generate crucial information which can be used to adopt appropriate measures reducing the aftermath of crashes. Many scientific studies developed different kinds of machine learning (ML) models for crash severity predictions. However, there is a need of the model using relevant data reasonably covering the world which can be customized for new datasets with limited training data. It will pave the path towards the development of a global solution to traffic crash severity prediction. This study aims to build the solution through the synergy between big data and deep learning. The initial step is to develop an international database of traffic crash severity consists of publicly available resources from relevant authorities and academic publications. Different deep learning models along with many learning and regularization approaches will be investigated to develop an international model which will address the bias and variance issues adequately. It is assumed that the deep learning model will provide a high-level understanding of the features. Therefore, the model is expected to perform well even with limited training datasets for new locations. It will also be compared with conventional machine learning models. The output of this research will mainly assist proper actions and remedies required to enhance traffic safety in the developing countries due to the limited availability of traffic crash data. Moreover, the output of this research can be utilized by researchers globally and international entities such as World Health Organization, United Nations.