Few-shot cross-city traffic injury severity prediction: a context-aware domain generalization meta-learning approach

Developing accurate traffic injury severity prediction models for cities with limited historical data remains a critical challenge. The primary obstacle stems from the significant spatial heterogeneity of traffic environments, where data-driven models trained on source domains often exhibit poor generalization to target cities with distinct road geometries and crash characteristics. Conventional transfer learning and standard meta-learning algorithms typically struggle to mitigate negative transfer, as they lack explicit mechanisms to decouple domain-specific noise from generalizable safety patterns. To address these limitations, this study proposes the Context-aware Domain Generalization Model-Agnostic Meta-Learning (CaDG-MAML) framework. First, a rigorous domain holdout strategy is incorporated into the meta-training loop, simulating the shift to unseen domains to compel the model to prioritize domain-invariant feature representations. Second, to resolve local heterogeneity, a Context-FiLM (Feature-wise Linear Modulation) module dynamically recalibrates task-specific decision boundaries by modulating intermediate feature maps using aggregated support-set statistics. The proposed framework is evaluated on UK STATS19 two-vehicle crashes (2019-2023) under a 2-way 5-shot setting, with four held-out target cities. Results show that CaDG-MAML achieves the best four-city average performance on AUC, Balanced Accuracy, F1-Severe, F1-Macro, and G-Mean among eight baseline models, with particularly stable advantages on severe-crash-oriented metrics. Paired statistical tests indicate that these gains are generally significant relative to the strongest baseline. Hyperparameter sensitivity and ablation analyses further support robustness and component complementarity. Dynamic SHAP analysis indicates that the framework preserves established risk-related feature groups while recalibrating feature weights across cities to inform tailored safety interventions.