Dos dados ao alerta: utilizando machine learning para prever e explicar a gravidade de acidentes em rodovias federais brasileiras

Abstract

Traffic accidents on federal highways represent a complex challenge for public safety and infrastructure management in Brazil. This study aims to develop an interactive web application based on Machine Learning to estimate the probability that a person involved in an accident will experience a severe outcome (serious injury or death) and to explain the factors associated with the prediction. The methodology was based on training a predictive model using the XGBoost algorithm, fed with historical accident data and information on the vehicle fleet. To ensure system transparency and avoid the opacity inherent in complex models (the “black-box” effect), the SHAP (SHapley Additive exPlanations) library was integrated to provide local explainability for the predictions. The solution was developed using a modular architecture, separating the web interface, built with Streamlit, from the services responsible for data loading, input processing, prediction generation, and result interpretation. The results demonstrate that the dashboard allows users to enter physical, temporal, and environmental characteristics of a road segment and, in real time, estimate the probability that a person involved in an accident will sustain a serious injury or die, while also displaying a graph detailing the contribution of each variable to the prediction. It is concluded that the tool functions as a predictive and interpretable decision-support system, with practical potential to assist agencies such as the Brazilian Federal Highway Police (PRF) and the National Department of Transport Infrastructure (DNIT) in the preventive allocation of enforcement resources and the prioritization of improvements to the road network.