Study of Risk Factors and Prediction of Road Accident Severity in Brazil

Abstract: This study seeks to advance the understanding of road accident severity by addressing underexplored aspects in the literature and analyzing road maintenance characteristics, driver profiles, traffic conditions, and environmental influences through machine learning models. The research draws on data from accidents on federal highways across Brazil in 2021 and 2022, covering all states and totaling 8,537 crashes. Variables include traffic volume, road infrastructure, driver profile characteristics, and environmental factors. The method comprises (1) integrating data sources to create a unified database, (2) transforming categorical variables into dummies and selecting relevant predictors, (3) splitting the data into training (80%) and testing (20%) sets, and (4) applying decision tree, random forest, gradient boosting, and zero-inflated logistic regression models to predict accident severity. Across all models, key variables - Annual Average Daily Traffic (AADT), driver age (notably 18–35 years), and vehicle age (vehicles over 10 years old) - emerged as significant predictors, underscoring the role of traffic exposure, driver characteristics, and vehicle condition in accident severity. Random forest and gradient boosting captured a broader range of variables, including environmental conditions, while decision trees and logistic regression models tended to assign less importance to factors such as vehicle type and weather conditions. Despite minimal differences in accuracy between models (random forest = 87.84%; gradient boosting = 87.99%; decision tree = 81.50%; zero-inflated logistic regression = 87.96%), the interpretability of decision tree and logistic regression models makes them especially valuable for identifying the relative importance of risk factors. These findings offer a comprehensive perspective on the factors influencing severe road accidents, providing a pathway for road managers to prioritize interventions based on a thorough assessment of risk factors. Study limitations include a lack of detailed highway geometry data, regulatory speed information, pavement macrotexture and friction data, and limitations in AADT accuracy. Future research should expand the sample and develop Safety Performance Functions (SPF) by incorporating non-accident data.