Modelagem preditiva da ocorrência de acidentes graves na BR-101 em Sergipe utilizando árvores de decisão

Abstract

This study aims to analyze the factors that most influence the survival of people involved in traffic accidents on BR-101, in the state of Sergipe, based on data provided by the Federal Highway Police (PRF). The dataset used contains detailed information about the accidents, including weather conditions, road conditions, type and number of lanes, time of accidents, type of collision and severity, among other relevant aspects. The analysis focuses on the classification of accidents according to the survival of victims, differentiating between “No Victims” and “With Victims” cases. The main objective is to identify how environmental and structural factors affect the probability of survival. Models such as CART (Classification and Regression Trees), Bagging and Random Forest were used. The results showed that, although the CART model achieved the highest accuracy of 0.676, it performed poorly in identifying serious accidents, with a recall of only 0.098. The balanced version of CART improved the recall by 0.329, but reduced the accuracy. Aggregation methods, such as Bagging and Random Forest, demonstrated a better ability to identify serious cases, with recalls above 0.57 and F1- Scores around 0.45, standing out as the most effective approaches to the problem. Based on this approach, the study aims to provide a detailed analysis of accidents on BR-101 in Sergipe, assisting public managers in implementing road safety policies and defining priority preventive measures to reduce the severity of accidents and increase highway safety.