Traffic collisions present a persistent public safety and congestion challenge in large urban environments such as Toronto. Existing studies often rely either on predictive machine learning models that lack downstream traffic interpretation or on traffic simulation frameworks that depend on static assumptions. This study introduces a novel integration of machine learning–based risk prediction with Cellular Automata simulation to dynamically model congestion propagation under high-risk conditions. This study addresses this gap by proposing an integrated framework combining exploratory spatial analysis, supervised machine learning, and an adaptive Cellular Automata (CA) traffic simulation to model collision risk and congestion under high-risk conditions. Using over 500,000 geo-referenced police-reported collision records from the City of Toronto, spatial analysis identifies high-risk zones, including the downtown core and Scarborough, and peak collision periods during morning and evening rush hours. Random Forest and XGBoost models are used to predict injury severity, achieving 83% accuracy, an F1-score of 0.79, and an AUROC of 0.86, outperforming linear baseline models. The predicted risk outputs are integrated as dynamic inputs into a Toronto-calibrated CA model to simulate congestion propagation under adverse weather, visibility, and behavioral conditions. The simulation reproduces observed congestion patterns with 88% spatial–temporal agreement. These results demonstrate the effectiveness of combining predictive analytics with simulation-based modeling to support data-driven urban traffic safety planning. The proposed framework supports proactive traffic safety interventions aligned with the City of Toronto’s Vision Zero objectives.
Boddepalli et al. (Wed,) studied this question.