The impact of non-functioning traffic signals on pedestrian injury severity: evidence from a nationwide interpretable machine-learning study (Poland, 2015-2024).

Pedestrians face disproportionate injury risk in traffic. Using nationwide SEWIK police records (2015-2024), I model five-level pedestrian injury severity with proportional-odds logistic regression, Random Forest and XGBoost under time-blocked, incident-aware cross-validation, and I interpret Random Forest predictions with SHAP while simulating repaired traffic signals. Random Forest attains the strongest ordinal agreement (quadratic weighted kappa 0.241 ± 0.042 across eight outer folds; bootstrap 95% CI 0.212-0.264), clearly outperforming naive baselines (majority-class and ordinal-median QWK ≈ 0), whereas XGBoost peaks on accuracy (0.556) and F1-weighted (0.496). SHAP highlights age, lighting, precise location in the road space and posted speed limit as dominant severity drivers; age ranked first and daylight second in seven of eight outer folds. Counterfactual signal restoration on n = 1,212 records with a non-functioning signal shifts predicted mass towards milder outcomes, most visibly 74 transitions from seriously to slightly injured (bootstrap 95% CI 59-92) and 31 from died within 30 days to seriously injured (CI 21-42). Predictive performance remains modest (macro ECE 0.115); estimates are prognostic and Poland-specific, and counterfactuals are associative rather than causal, but the workflow illustrates how registry analytics can inform maintenance prioritization for life-saving infrastructure.