Dual-Pathway Analysis of Severe Occupational Accident Outcomes Using Narrative-Derived Scenario Variables and Explainable Machine Learning

Occupational accident severity is often modeled as a single ordered outcome, potentially obscuring differences between fatal and severe non-fatal injuries. This study develops a dual-pathway framework that separately analyzes fatal versus non-fatal outcomes and severe versus minor injuries among non-fatal cases. Using 48,909 accident records from the CSDataset, structured attributes were integrated with narrative-derived scenario-exposure and work-activity variables obtained through rule-constrained LLM coding; expert validation yielded Cohen’s κ = 0.742. Logistic and Firth logistic regressions were combined with machine learning models, SHAP analysis, and scenario–activity co-occurrence analysis. The two pathways shared several explanatory variables but differed in association direction and magnitude. In the fatality pathway, vehicle and mobile-equipment exposure (OR = 1.813, 95% CI: 1.668–1.971) and workplace violence (OR = 5.184, 95% CI: 4.190–6.415) showed positive associations. Among non-fatal cases, driving operation showed the strongest positive activity-level association with severe injury (OR = 2.188, 95% CI: 1.715–2.793). Co-occurrence analysis showed that frequent combinations were not necessarily outcome-enriched; vehicle and mobile-equipment exposure combined with driving operation was enriched in both pathways. These findings support pathway-specific interpretation and work-context-oriented safety management and prevention.