A Multi‐Task ConvoBiLSTM Model With Self‐Attention for Concurrent Forecasting of Traffic Accident Risk and Severity

In this study, we introduced a multi‐task deep learning framework that concurrently forecasts traffic accident risk and severity by integrating convolutional neural networks (CNNs), bidirectional long short‐term memory (BiLSTM) units, and a self‐attention mechanism. Unlike conventional single‐task approaches, our model leverages shared spatiotemporal representations to capture complex patterns in traffic data, thereby enhancing both predictive accuracy and generalizability. Evaluations on large‐scale datasets from New York City and Chicago demonstrate that our approach achieves high accuracy (up to 92% for accident risk and 89% for severity) and remains robust across diverse urban contexts. Moreover, an enhanced SHAP‐based interpretability module provides granular insights into the influence of both observable and latent factors, such as driver behaviour or road surface conditions, on prediction outcomes. The self‐attention mechanism further mitigates unobserved heterogeneity by highlighting critical time steps and feature interactions. With competitive real‐time performance and throughput, our framework offers a practical solution for dynamic traffic safety applications. Future work will focus on extending evaluations to broader urban settings and integrating latent variable models to better quantify unobserved influences, ultimately advancing the development of safer, more efficient transportation systems.