Predicting Crash Severity using Naturalistic Driving Data and Neural Networks

Abstract This study leverages artificial intelligence (AI) to predict crash severity using naturalistic driving data (NDD) from the SHRP-2 dataset. It employs logistic regression and Recursive Feature Elimination (RFE) for feature selection, SHapley Additive exPlanations (SHAP) for model interpretation, and Feedforward Neural Networks (FNN) for prediction. The FNN models achieved high accuracy, correctly classifying severe crashes at 93.20% and moderate crashes at 94.98%. The Synthetic Minority Oversampling Technique (SMOTE) improved predictive performance for severe crashes, which were less frequent in the dataset. SHAP analysis identified near-miss events and driver responsibility indicators as key predictors of crash severity. The study underscores the importance of AI-driven predictive analytics in traffic safety, emphasizing the need for targeted interventions to mitigate crash risks. These insights can guide policymakers in developing effective strategies to enhance road safety and integrate AI technologies into traffic management systems.