This study aims to address the limitations of lagging indicators and static risk assessments by developing an automated framework to model the dynamic nature of construction accident causation. The objective is to transition safety management from reactive compliance to predictive risk mitigation by forecasting how hazards propagate across construction sites over time.
The framework uses construction incident reports from the Occupational Safety and Health Administration database. The methodology begins by applying natural language processing and semantic clustering to define a validated hazard taxonomy from unstructured narratives. Next, a temporal graph is constructed to model hazard co-occurrences within a 30-day metropolitan window. Finally, edge-aware Graph Neural Networks, including GraphSAGE, graph attention network, Graph Convolutional Network and Graph Isomorphism Network, are used to formulate hazard forecasting as a link prediction task.
The GraphSAGE architecture equipped with convolutional block attention module attention yielded the highest predictive performance (mean reciprocal rank = 0.537), outperforming baseline models. Temporal features were identified as the primary driver of predictability, outweighing static structural associations. The model successfully mapped high-probability hazard chains, such as the progression from demolition to structural and struck-by incidents.
The predictive framework enables safety managers, regulators and planners to anticipate imminent risks and deploy targeted interventions, optimizing resource allocation and site inspection schedules before sequential accidents occur.
This research provides an end-to-end framework that transforms unstructured text into a dynamic temporal graph. It offers a data-driven tool for forecasting regional hazard pathways, empirically validating theories that treat safety risk as a dynamic property of temporal alignment.
