Spatiotemporal grid-based crash prediction—application of a transparent deep hybrid modeling framework

Traffic crashes are one of the significant causes of death worldwide, and the prediction of this event is complicated due to many contributing factors. This study used spatial, temporal, and spatiotemporal information to predict crashes in Chicago at 1 km grid levels. A Deep Hybrid Network (DHN) was developed by exploiting inherent unique characteristics of Convolution Neural Network (CNN), Long Short-term Memory (LSTM), and Deep Neural Network (DNN). The hyperparameters of the models were obtained through the Bayesian optimization algorithm. The proposed modeling framework investigated the feature importance, the spatial heterogeneity of predictions, the worst-performing spatial grids, and the spatial distribution of features pertinent to model performance. These analyses transform the proposed DHN into an interpretable and transparent model. The DHN model was compared with Logistic Regression (LR), DNN, CNN, LSTM, and bidirectional LSTM, and it outperformed the baseline models with an accuracy of 0.72, recall of 0.70, false alarm rate of 0.28, and AUC of 0.79. The top three essential features were time, weather, and taxi trips, consecutively. The grid-level distribution of prediction performance investigations revealed a consistent performance of all deep learning models in terms of failed grids (i.e., AUC is 0.5 or less). It was revealed that DHN has the fewest failed grids (i.e., 18 failed from 710 grids) among the experimented models. According to the district level analysis, the O’hare airport area and the central district had the fewest number of failed grids for all methods, while the far south district had the highest number of failed grids. In addition, it was observed that the passed grid had a higher average feature density than the failed grid.