Development of Grid-based spatiotemporal crash prediction model

Abstract: According to the World Health Organization (WHO), millions of lives are lost in traffic road crashes each year. Predicting these traffic crashes poses a significant challenge due to the complex interplay of temporal, spatial, and spatiotemporal factors. In this study, we focus on predicting traffic crashes in Nashville by leveraging temporal, spatial, and spatiotemporal data collected at 5 km grid intervals. To address this challenge, we propose a novel Modified Hybrid Model (MHM) that draws inspiration from LSTM (Long Short-Term Memory), CNN (Convolutional Neural Network), and XGBoost techniques. The MHM is designed to capture the intricate patterns and relationships within the data. We also employ Meta Learning to fine-tune the model's hyperparameters, ensuring optimal performance. In addition to prediction, our research explores various aspects of traffic crash analysis. We investigate outlier detection, model calibration, feature importance, identification of worst-performing spatial grids, spatial distribution of features, and spatial heterogeneity of predictions. This comprehensive approach provides valuable insights into the factors contributing to traffic crashes in Nashville. To assess the effectiveness of our MHM model, we conducted a comparative analysis with several widely used models, including XGBoost, LSTM, CNN, Logistic Regression (LR), and bidirectional LSTM. Our findings demonstrate the superiority of the MHM model in predicting traffic crashes, highlighting its potential for improving road safety and saving lives. By leveraging the power of hybrid models and advanced techniques, we aim to contribute to the development of more accurate and effective strategies for road safety and accident prevention.