Feature Selection and BPNN Optimization Using RIME Algorithm for Freeway Crash Clearance Time Prediction

ABSTRACT Accurate prediction of crash clearance time is crucial for enhancing emergency response efficiency and promoting road safety. Traditionally, such predictions have relied on statistical models, which are often constrained by limited capacity to capture complex nonlinear relationships and dependence on strict distributional assumptions. Therefore, recent studies have increasingly employed machine learning methods to enhance modelling flexibility and prediction accuracy. However, these approaches commonly face two key limitations: limited interpretability and inadequate attention to feature selection. To overcome these challenges, this study introduces a novel model—rime optimization algorithm–backpropagation neural network–MF (RIME‐BPNN‐MF)—which improves the performance of the backpropagation neural network (BPNN) by leveraging the RIME optimization algorithm for feature selection and model parameter tuning. The model is tested using five years of freeway crash data from South Korea. According to the distribution of crash clearance time, the dataset is divided into two categories using a 34‐min threshold. Comparative experiments showed that the proposed model outperformed the baseline models in predictive performance for both short‐ and long‐clearance times. To improve interpretability, the Shapley additive explanation (SHAP) is applied, revealing that vehicle damage severity is the most critical factor for predicting short clearance time, whereas crash severity plays the most significant role in long clearance time. This study is the first to apply an optimization algorithm for feature selection in crash clearance time prediction, providing new insights into effective feature selection strategies in this field. The outcomes of this study may be of interest to traffic agencies and operators for response planning and duration‐specific resource dispatch.