An explainable multi-task deep learning framework for crash severity prediction using multi-source data.

Traffic accidents pose significant global challenges, causing substantial injuries, fatalities, and economic losses. Current research predominantly focuses on single-prediction objectives (e.g., fatality prediction) while neglecting property damage assessments and critical interactions between prediction tasks. Although neural networks demonstrate superior predictive capabilities, their application in traffic safety analysis remains constrained by inherent limitations in causal interpretability, coupled with challenges posed by data imbalance, heterogeneity, and complexity in crash datasets. This study proposes an interpretable multi-task learning framework (Adv MT-DNN) that synergistically integrates an enhanced deep neural network with post-hoc explanation methods for comprehensive crash severity prediction. Our dual-focused approach addresses multiple prediction targets (including fatalities, severe injuries, and property damage). It provides granular insights into contributing factors through SHAP-based feature importance rankings and interaction analysis. Validated using four-year (2018-2021) multi-source traffic data from China, the framework demonstrates significant improvements in prediction accuracy compared to baselines. Nonparametric estimation of the top-8 critical factors (e.g., blood alcohol content, collision type, and accident occurrence period) confirms statistically significant associations with crash severity. The explicit interpretation mechanism bridges the critical gap between predictive performance and model interpretability in traffic safety analytics, providing engineering-relevant insights. This research establishes a robust methodological foundation for developing data-driven road safety policies and intelligent transportation systems, particularly in developing countries with complex traffic ecosystems.