This study develops a data-driven severity-classification approach for maritime construction, relating equipment activity from incident narratives to localized hydro-meteorological conditions to complement static safety systems.
The research analyzed 1,139 maritime construction incidents from OSHA databases (2015–2025). A rule-based Natural Language Processing module classified unstructured narratives into 15 distinct equipment categories. Additionally, an interface to a historical weather API reconstructed micro-climate conditions at the incident locations. Sixteen machine-learning algorithms were comprehensively compared for injury severity classification using chronological hold-outs, stratified cross-validation and feature-ablation.
A Logistic Regression classifier provided a strong balance of discrimination and interpretability, achieving a cross-validated AUC of 89.2% and a hold-out AUC of 95.6% for post-incident classification based on narratives, while a pre-incident configuration using only prior operational factors achieved an AUC of 68.7%, with an F1-score of 95.4% and a Brier score of 0.062. The predictive signal originates primarily in the incident narrative; weather and employer history contributed modestly. A sensitivity analysis confirmed performance was robust to employer-history exclusion, and a localized wind-speed association near 30 km/h was cautiously identified.
The approach offers a conceptual triage aid for site superintendents. If integrated into Construction Safety Management Systems, this logic could prioritize hazard reviews and inform daily planning, pending operational validation.
The study integrates unstructured narratives with quantitative geospatial weather metrics in maritime construction. It reports a highly transparent classifier as an alternative to opaque models, offering an applied integration of established methods for safety-critical risk analysis.
