Comparative evaluation of deep learning and traditional models for predicting traffic accident severity in Saudi Arabia
Fayez Alanazi, Ibrahim Khalil Umar, Ahmed M. Yosri, Mohamed Ahmed Okail
Abstract
Road traffic accidents are one of the leading death causes around the globe, claiming millions of lives every year. Predicting traffic accident severity is essential for road users' safety and accident prevention. Artificial neural network (ANN), Boosted trees (BRT), Support vector machine (SVM), Naïve Bayes (NVB), and logistic regression (LGR) were employed for predicting fatal accidents in 14 cities in the Eastern Province of the Kingdom of Saudi Arabia using accident data from the year 2018-2022. The accident data was classified into fatal and injury accidents. A total of 9,548 accident cases involving 17,100 vehicles resulting in 2,527 fatalities and 8,020 injuries during this period, with 28% of the cases occurring in Al-Ahsa. The ANN model outperformed all five models with an accuracy = 99.91%, sensitivity = 99.94%, specificity = 99.8%, G-mean = 99.87%, and AUC = 99.92%. The ANN could improve the performance of LGR by up to 13.60% in the validation phase. For understanding the impact of each of the input parameters, three different relevance-ranking algorithms (maximum relevance minimum redundancy, Kruskal Wallis and Chi-square) were used prior to the development of the models and the result shows the number of people involved and the number of people injured as the major factors increasing the severity of road traffic accidents.