Highway to... Determining Fatal Outcomes in Traffic Accidents Based on Police Reports

• Arthur M. P. Gabardo UFSC
• Guilherme A. A. Schünemann UFSC
• Pablo A. Jaskowiak UFSC
• Benjamin G. Moreira UFSC
• Ricardo J. Pfitscher UFSC

Resumo

Brazil faces significant traffic safety challenges with its vast territory and one of the world’s largest road networks. Road traffic accidents, particularly on federal highways, remain a leading cause of death in the country, with serious economic and social consequences. This work presents a case study of three machine learning methods—Random Forest (RF), k-Nearest Neighbors (kNN), and Multilayer Perceptron (MLP)—for classifying the severity of traffic accidents in the Brazilian southern region. Using an open dataset from the Brazilian Federal Highway Police (PRF) covering the years 2021 to 2024, extensive preprocessing was carried out, including categorical variable encoding, feature selection, and the application of the SMOTE technique to address class imbalance. Model performance was assessed through statistical metrics such as specificity, F1-score, and AUC-ROC. The results show that RF and kNN (with SMOTE) achieved the best performance in predicting fatal accidents, both with AUC-ROC of 0.99. In addition to an in-depth model evaluation, this study presents a post-hoc analysis of feature importance and contributions through Shapley Additive Explanations (SHAP) for the best performing model, in order to support knowledge discovery and highlight the most influential factors associated with fatalities.