Supervised clustering using SOM for severity-based pattern detection in urban traffic crashes

Abstract

Urban traffic crashes remain a critical public health challenge, particularly for vulnerable road users. This study introduces a novel data-driven methodology to support urban road safety planning by identifying well-defined, interpretable crash typologies associated with fatal or serious injuries. The proposed framework relies on a supervised clustering strategy that integrates SHAP (SHapley Additive exPlanations) values with Self-Organizing Maps (SOM). Applied to urban crash data from Barcelona (2017-2019), the approach uncovers ten distinct and interpretable crash typologies, capturing high-risk scenarios such as speed-related nighttime collisions and pedestrian-heavy vehicle conflicts, as well as less explored patterns including two-wheeler falls and bicyclemotorcycle interactions. By combining SHAP-based explanations with topology-preserving neural mapping, the SOM framework reveals subtle gradations of risk, preserves neighborhood relationships among crash profiles, and enhances subgroup detection and interpretability beyond traditional unsupervised clustering methods and standard eXplainable Artificial Intelligence (xAI) summaries. These results underscore the potential of SOM-based supervised clustering to inform targeted, data-driven safety interventions. More broadly, the study advances methodological research on supervised clustering and offers a transferable tool for detecting high-dimensional risk patterns in urban safety analysis and other applied domains.