Motorcycle Safety Management through Injury Severity Monitoring: Evidence from Poland
Purpose: This paper examines how national police-reported crash data can support road traffic safety (RTS) management by enabling systematic monitoring of motorcycle injury severity and helping authorities prioritise safety risks, without treating predictive scores as a substitute for clinical triage. Design/Methodology/Approach: The study analyses 83,473 motorcycle riders (excluding moped riders) recorded in Poland’s SEWIK police crash registry over the period 2015–2025. Injury severity, classified into five ordered categories, is modelled using the CatBoost multi-class gradient-boosting algorithm, while SHAP (Shapley Additive Explanations) is used to interpret feature associations. A data-quality audit identifies registry-process artefacts, including vehicle production year, protective-equipment status and calendar year. These variables are excluded because their inclusion substantially undermines out-of-time predictive performance around a 2021 recording discontinuity. Model performance is then assessed using walk-forward validation. Findings: Following feature ablation, ordinal agreement remains modest, with a mean quadratic weighted kappa of approximately 0.25 across walk-forward validation. Crash type, posted speed limit, licensing status and impairment exhibit the strongest associations with predicted injury severity. The results highlight the importance of addressing temporal discontinuities and administrative recording artefacts when developing predictive monitoring tools from crash data. Practical implications: Road safety authorities can use the proposed pipeline as a supplementary decision-support tool to monitor injury-severity patterns, identify risk factors warranting further investigation and inform the prioritisation of motorcycle safety assessments. The findings may help practitioners structure dashboard indicators around crash characteristics, speed limits, licensing and impairment, while incorporating routine data-quality checks and temporal validation. Given the modest predictive agreement, outputs should support—not determine—resource allocation and should not be interpreted as individual-level risk assessments or evidence of intervention effectiveness. Originality/value: The paper contributes to RTS management by integrating explainable machine learning, temporal validation and explicit data-quality governance into a transparent pipeline for monitoring motorcycle injury severity using national police crash data. It provides a framework for responsible use of predictive analytics in safety management while distinguishing statistical associations from causal effects.