Predicting Road Accident Severity Using Traffic, Accident and Network Data

Road traffic accidents present significant public safety and economic challenges, necessitating accurate severity prediction for targeted safety interventions. Traditional models rely on historical crash data but often overlook real-time traffic conditions and road geometry. This study proposes a machine learning framework integrating traffic flow data, road network attributes, and accident records to enhance severity classification. A key focus is assessing the impact of excluding post-accident features to ensure predictions rely solely on pre-crash conditions. A web-based application automates network extraction, traffic assignment, and data integration, making the framework scalable across urban environments. The approach employs XGBoost for severity prediction and SHAP for feature importance analysis, validated using Rome as a case study. Results highlight traffic flow, speed variance, and road design as critical severity determinants.