Shedding light on crash injury severity: a Bayesian-optimized machine learning investigation of visibility conditions

Visibility conditions play a critical role in road traffic crash outcomes, yet limited research has examined how combinations of ambient lighting and street lighting are jointly associated with injury severity. This study examines the association between visibility conditions and traffic crash injury severity using a database of 17,840 person-level records of casualties in Bahrain. In addition to demographic, road, environmental and crash-related factors, five visibility categories combining ambient and street lighting conditions were explored. Five machine learning models were developed using nested cross-validation with Bayesian hyperparameter tuning (Optuna). CatBoost ranked highest with modest discriminative ability (ROC-AUC = 0.691 ± 0.007), though the gradient-boosting models performed comparably. SHapley Additive exPlanations (SHAP) analysis identified Year, Gender, Person Involved, and the temporal Month and Day components as the most influential predictors, followed by Nationality and Cause Type. Among visibility categories, Night-Street Light Lit exhibited the strongest associations and interactions, most notably a weaker association with severity for female casualties, alongside weaker seasonal and passenger-related associations. The unlit condition displayed minimal interactions. These results point to a dominant additive role of visibility conditions alongside selective context-dependent associations for certain road-user and seasonal subgroups, supporting targeted road safety interventions that incorporate lighting factors.