Coal Mine Accidents in China and the United States, 2011–2025: Harmonized Spatio-Temporal Patterns, Key Risk Factors, and Interpretable Machine Learning Severity Prediction from Public Accident Data

The two largest coal producers, China and the United States, operate markedly different accident-reporting regimes, yet comparable cross-country evidence is scarce. We harmonized 47,986 U.S. Mine Safety and Health Administration (MSHA) Part 50 records for 2011–2025, documenting reportable injuries, occupational illnesses and accidents, with Chinese official statistics and a published 532-report structured dataset. Because the archives report different objects, and China reports fatalities per million tonnes while U.S. reporting is event-based, we placed both on a common production-normalized basis, tested correlates of fatal outcomes with chi-square statistics and Cramér’s V and benchmarked four classifiers under explicit class-imbalance controls with SHAP attribution and leakage-controlled narrative text. China’s rate fell from 0.564 (2011) to 0.045 (2025) while the U.S. rate stayed flat (0.020 to 0.017), narrowing the gap from 28-fold to under 3-fold. U.S. fatalities concentrated in powered haulage (31.2%) and machinery (25.9%); Chinese deaths were dominated by gas explosions and outbursts (20.8 and 9.6 per event). Scrubbed narratives raised cross-validated ROC–AUC from 0.892 to 0.966 and held-out PR–AUC from 0.049 to 0.252; a sensitivity analysis confirmed the composition findings under the Chinese dataset’s severity bias. The framework supports screening-oriented risk triage of newly reported events, not automated adjudication.