Urban Road Crash Severity Prediction Using Explainable Machine Learning and Spatial Hotspot Analysis: Evidence from Loja, Ecuador

Administrative road crash records in intermediate cities may support safety planning but often contain incomplete outcomes and inconsistent coordinates. We analyzed 3349 events recorded in Loja, Ecuador, during 2020–2026 (2026 partial). Severity was reconstructed from injury and fatality counts; crashes involving either were operationally classified as severe. Penalized logistic regression, Random Forest, and XGBoost were trained on 2020–2024 records (n = 2703) and tested on 2025 (n = 533), using a conservative predictor specification to reduce administrative outcome leakage. After correcting invalid coordinates, XGBoost achieved balanced accuracy of 0.705 and ROC-AUC of 0.766; Random Forest achieved 0.688 and 0.760 and was retained for permutation-based interpretation. Joint domain permutations identified crash typology, vulnerable-road-user indicators, and spatial context as predictive sources of information, not causal effects. Among 3311 georeferenced records, the original 250 m grid assigned 3309 and identified 50 very-high-priority cells. Sensitivity analyses examined coordinate representation, grid resolution, and two boundary-located observations. The workflow provides an auditable approach to retrospective classification and spatial screening, but cannot establish prospective crash risk without further exposure and external validation.