A Dimensionality-Reduced XAI Framework for Roundabout Crash Severity Insights

Roundabout safety isn't one-size-fits-all: The study identified four distinct crash patterns in Ohio roundabouts (2017-2021), each with different severity drivers, meaning that effective safety interventions must be tailored to specific conditions rather than applying uniform solutions across all roundabouts.

Darkness and wet conditions dramatically increase injury risk: The research found that crashes are significantly more severe when darkness, wet road surfaces, and higher posted speeds occur together with fixed-object or angle collisions, suggesting that targeted safety measures (like improved lighting or speed management) could meaningfully reduce harm in these high-risk scenarios.

Different crash locations require different safety strategies: The analysis revealed that crashes have distinct mechanisms depending on location—fail-to-yield and gap acceptance issues at entries, improper maneuvers in multi-lane areas, and rear-end collisions during slow-downs—enabling transportation agencies to design location-specific countermeasures rather than generic fixes.

Explainable AI can make safety decisions more transparent and accountable: By using interpretable machine learning methods (SHAP and Cluster Correspondence Analysis), the framework produces audit-ready explanations for why certain crashes are predicted to be severe, supporting public accountability and helping officials justify safety investments to communities.

This approach bridges the gap between data patterns and real-world action: The two-step workflow connects statistical pattern discovery with case-level explanations, providing a practical template that transportation agencies can use for site screening, selecting appropriate safety countermeasures, and communicating decisions to stakeholders.