A systematic mapping review of explainable machine and deep learning approaches for road traffic crash severity prediction: A decade-long review (2014-2024)

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Identifying the severity of road accidents is critical for effective traffic management and emergency response. Machine learning has been increasingly applied to predict crash severity, but the lack of transparency in many black-box models poses a significant challenge, especially when human lives are at stake. Governments worldwide are now mandating explainable artificial intelligence in critical applications, emphasizing the need for transparent decision-making processes. In our review paper, we address this issue by consolidating and categorizing the various machine and deep learning techniques used in road traffic crash severity prediction. We systematically analyzed 237 articles published between 2014 and 2024 in the SCOPUS and Web of Science databases using a combination of specific keywords, categorizing them based on their use of explainable techniques, including model-agnostic, model-specific, or feature-importance, and ablation techniques. Among the 237 studies, 51% used only ML models, 16% used only DL models, and 33% compared both ML and DL approaches. Our analysis revealed that approximately 64% of the reviewed studies employed ablation methods for feature selection. Random Forest emerged as the most frequently adopted machine learning model, whereas feature-importance mechanisms, particularly in tree-based ensemble models, remained the dominant explainability approach. Additionally, researchers employed model-agnostic methods to explain deep learning-based models. SHAP emerged as the most widely adopted model-agnostic explainability technique and exhibited a clear increase in adoption over recent years. A small number of studies delve into model-specific explanations. Overall, model-agnostic techniques were found to be more widely adopted than model-specific explainability approaches for interpreting complex predictive models. The integration of explainable crash prediction models with emergency response systems shows promise for this application. This taxonomy provides a structured resource for researchers and offers insights for policymakers, transportation experts, and emergency response professionals, thereby supporting efforts to improve road safety and the quality of emergency services.