Harnessing Machine Learning to Enhance Global Road Safety: A Comprehensive Review

Abstract

As global urbanization accelerates, road safety remains a pressing concern, underscored by escalating traffic accidents and fatalities. Road Traffic Injuries (RTI) have become the eighth leading cause of death worldwide. The article delves deep into the potential of machine learning in predicting traffic accidents, their severity, and causal factors. This study comprehensively evaluates machine learning models on traffic accident records sourced from the Addis Ababa City Police Department. Comprising 12,316 records with 15 features, the dataset underwent preprocessing techniques, specifically Synthetic Minority Over-sampling Technique (SMOTE) and Min-Max scaling. Five algorithms – Random Forest (RF), Gaussian Naive Bayes, CatBoostClassifier, LightGBM, and XGBoost – were tested for their prediction accuracy. The findings spotlight the dominance of the RF model, achieving a peak accuracy of 92.2% post-SMOTE and Min-Max application. A comparative analysis with existing literature showed that while RF is a recurrently effective model across various datasets, data preprocessing and model suitability to specific datasets is paramount. This study underscores the potential of machine learning in traffic accident analysis and the nuanced choices researchers must make for optimal outcomes.