Deep learning for classifying and predicting risk factors in retrospective construction fatality reports

Abstract

Despite advances in safety protocols, fatal incidents continue to occur in the construction industry. The National Institute for Occupational Safety and Health (NIOSH), Fatality Assessment and Control Evaluation (FACE) program in the U.S. documents hundreds of fatalities; however, the unstructured narrative nature of these reports constrains systematic analysis. This pilot study applies deep learning to classify and interpret fatal construction incidents from 265 NIOSH FACE reports, primarily using narrative text, with structured attributes supporting annotation and label development. The approach involved curating and labeling incident narratives, fine-tuning transformer-based models for supervised classification, and evaluating performance across four targets: Incident Type, Project Type, Incident Causation, and Temporary Structure Type. Results indicated strong performance for Incident Type (accuracy = 0.981 ± 0.019; macro-F1 = 0.933 ± 0.108), moderate performance for Temporary Structure Type (accuracy = 0.830 ± 0.013), and mixed outcomes for Incident Causation and Project Type due to semantic overlap and class imbalance. Overall, the findings demonstrate the feasibility of converting narrative fatality data into predictive insights and support the development of scalable, data-driven frameworks for improving construction safety research and intervention.