To identify the key determinants of urban road-traffic accident severity and their interaction effects, we analyze 6,968 accidents handled under the general procedure in Shenzhen during 2016–2019 and develop a Bayesian-network framework that integrates CatBoost–SHAP feature screening, K2 structure learning, and maximum likelihood estimation (MLE). Coupling backward (inverse) inference with scenario analysis, we quantify severity migration and risk amplification using the probability difference (ΔPc) and the amplification index (Rc). The results indicate that vehicle type, road lighting, motor-vehicle compliance/inspection status (e.g., unresolved violation or expired inspection), and vehicle technical condition are principal contributors. Scenario analyses constructed from these determinants reveal systematic increases on the more-severe end (ΔPc >0 and Rc >1) driven by multi-factor coupling. Under stratified five-fold cross-validation, the model attains an overall accuracy of 80.37%, with AUCs of 0.839, 0.815, and 0.774 for the three severity classes (Damage,Injury, Fatal). The framework balances predictive performance with interpretable inter-factor dependencies, elucidates the risk-amplification mechanism of multi-factor interactions, and provides quantitative evidence to support scenario-based joint interventions and urban traffic-management measures.