A state-of-the-art review of injury severity analysis in traffic crashes: toward a generalized modeling framework

Abstract

The modeling of injury severity (INJ-S) in road traffic crashes has evolved significantly with the advent of machine learning (ML) and hybrid data-driven approaches. This review provides a comprehensive synthesis of over 70 recent studies across diverse geographic regions and crash contexts, focusing on the integration of ML and advanced econometric techniques for predicting injury outcomes. The article examines a wide range of datasets, severity classifications, and methodological innovations, highlighting how ensemble methods, deep neural networks, and random parameter models enhance predictive performance and interpretability. It also emphasizes the critical role of spatial and temporal factors in shaping INJ-S, advocating for context-aware models that account for regional and time-based variability. Key findings highlight the persistent influence of behavioral, environmental, vehicular, and infrastructural factors on crash severity, as well as the increasing importance of preprocessing, feature engineering, and post-modeling interpretability tools, such as SHapley Additive exPlanations (SHAP) and global sensitivity analysis (GSA). A generalized analytical framework is proposed to guide future research, spanning data acquisition, model development, and policy-relevant interpretation. By bridging transportation engineering, data science, and public health perspectives, this study lays the foundation for scalable, interpretable, and geographically inclusive INJ-S modeling aimed at advancing global road safety.