Interpretable spatiotemporal traffic crash risk prediction using DMD-based graph neural networks

Spatiotemporal prediction of urban traffic crashes is an important basis for proactive safety management, yet many existing models have difficulty capturing both temporal dynamics and spatial dependence in an interpretable way. This study develops a hybrid framework that integrates Hankel-based Dynamic Mode Decomposition (Hankel-DMD) with a spatiotemporal graph neural network (STGNN) to predict short-term neighborhood-level crash counts. Daily crash records from 2019 to 2021 for 78 neighborhoods in Denver, USA are aggregated into a neighborhood-day matrix. Hankel-DMD is applied to this matrix to extract low-rank spatiotemporal modes that describe dominant trends and recurrent fluctuations. A graph neural network defined on a distance-correlation-based neighborhood graph then learns nonlinear residuals that correct the linear DMD prior and transmit information along the urban network. The proposed model is evaluated in a multi-step prediction setting with horizons of 1, 3, 5, and 7 days, and is compared with statistical time-series models, tree-based machine learning models, a pure Hankel-DMD model, and deep learning baselines including a STGNN. Across all horizons, the hybrid model achieves the lowest mean absolute error and root mean squared error, with improvements of about 17 to 30% in mean absolute error and 13 to 24% in root mean squared error relative to the best deep learning benchmark. Performance gains are consistent across high-, medium-, and low-risk neighborhood groups. Hankel-DMD eigenvalues and spatial modes reveal stable temporal and spatial structures in 2019 and 2021, together with clear deviations in 2020 associated with disrupted mobility patterns. These results show that dynamics-informed graph learning can provide accurate and interpretable crash risk forecasts at the neighborhood scale and can support targeted urban safety interventions.