Analysis of Motorcycle Accident Injury Severity and Performance Comparison of Machine Learning Algorithms

Abstract

Motorcycle road traffic accidents represent a big concern, as vulnerable road users account for more than half of all road deaths worldwide. The estimation of the influential factors associated with the increase of injury severity of motorcyclists involved in a road accident is of extreme importance as it provides a necessary basis for the development of an appropriate and targeted sustainable prevention plan for improving road safety. This study adopted the decision tree, ordered logistic regression (LR), random forest (RF), gradient boosting, extreme gradient boosting, k-nearest neighbor, and support vector machine methods to predict the injury severity outcome of motorcycle accidents. All the developed models were compared with six different performance metrics. A 10-year (2010–2019) dataset with motorcycle accidents that occurred in Portugal was used for analysis. As usual in traffic accidents datasets, this dataset is class unbalanced which was dealt with by under-sampling. The developed models made it possible to determine the factors associated with the increase of injury severity of motorcyclists involved in road accidents. The interpretation of each factor is based on the Shapley additive explanations values. The RF and LR models (developed with the balanced dataset) outperformed the other models. Risk factors associated with alcohol consumption, road type, road conditions, location, motorcycle age, rider’s gender, and when the accident occurs were estimated. This study provides a suitable framework analysis to build a proper predictive model, allowing researchers and practitioners to evaluate more accurately the risk factors of motorcycle injury severity.