Machine learning for traffic crash injury severity modelling: a systematic review integrating text mining analysis

This is the authors' abstract. We don't add key points for this paper.

Traffic crash injury severity modelling is critical for designing targeted road traffic safety interventions and management strategies. Existing reviews have the following limitations. First, they lack a systematic analysis of the complete machine learning workflow. Second, they do not fully utilize the valuable textual information such as abstracts through text mining. Third, current studies lack a unified standard for defining traffic crash injury severity, and existing reviews have not yet systematically summarized the relevant definitions. To bridge these gaps, this study conducts a systematic review that integrates a comprehensive examination of the machine learning pipeline with a text mining analysis of academic literature. Following the PRISMA guidelines and the predefined eligibility criteria, 306 studies were ultimately included for in-depth analysis. The review synthesizes definitions of injury severity and systematically examines data sources, influencing factors, predictive models, evaluation metrics, and key challenges such as class imbalance and model interpretability. From the text mining perspective, keyword frequency and burst analysis, co-occurrence network analysis, and research temporal evolution tracking were employed to reveal shifting research hotspots. Several key conclusions are obtained from this review. Notably, only approximately 37.6% of the reviewed studies addressed the class imbalance problem using techniques such as resampling, ensemble learning and loss functions. Similarly, approximately 61.8% of the studies explicitly incorporated model interpretability by employing methods such as LIME, SHAP, LSA, PDPs, GSA, and ICE. Furthermore, the field’s focus has progressively shifted from traditional statistical models toward advanced and explainable machine learning approaches. This review provides a comprehensive overview to help researchers quickly grasp key advancements and dynamics in the field.