Explainable Machine Learning for Cyclist Injury Severity in Bicycle–Vehicle Crashes in Poland: Association Patterns and Implications for Sustainable Road Safety
Road safety is a prerequisite for sustainable mobility, yet cyclists remain disproportionately exposed to severe outcomes in mixed traffic. Using police-reported bicycle–vehicle crashes from the national SEWIK registry in Poland (152,567 cyclist-involved records; 2015–2024), this study modeled five ordered injury-severity classes with a CatBoost gradient-boosting classifier, evaluated performance with quadratic weighted kappa and complementary class-sensitive metrics under extreme imbalance (including benchmark comparisons and calendar-based walk-forward stress tests), and interpreted predictions with SHAP to summarize transparent, feature-level association patterns. The results indicate modest overall ordinal discrimination (hold-out QWK ≈ 0.20), while highlighting elevated recall for rare fatal outcomes together with low precision, implying a substantial false-positive trade-off if outputs were used as deterministic classifiers. Global and local explanations point to stronger associations for cyclist age, shorter offender licensure tenure (a registry proxy for experience-related factors), regional context, and built-up versus non-built-up settings consistent with higher kinetic-energy environments; these variables should be interpreted cautiously because registry data are observational and omit key exposures (e.g., measured impact speed and cycling volume). Overall, the study contributes a nationwide, explainable severity-profiling workflow for prioritizing cyclist protection: combining benchmarked ML, multi-metric reporting, and XAI diagnostics can support monitoring and evaluation of speed management, infrastructure, and licensing-system improvements—without overstating causal effects from administrative records alone.