Association rule mining for proactive hazard identification in construction work activities

Construction remains one of the world’s highest-risk sectors, contributing a large share of occupational accidents. To improve proactive hazard identification, this study integrates Association Rule Mining (ARM) with the Cross-Industry Standard Process for Data Mining (CRISP-DM) to develop a structured, data-driven framework for safety learning. Although ARM has been used to uncover hidden relationships among accident factors, activity-specific hazard identification remains less. This study applies ARM to concrete elevated roadway superstructure construction in Indonesia using 131 documented accident cases. Four activity–hazard association rules were derived. The strongest rule showed that lifting box-girder segments was strongly associated with material damage, with support of 20.6%, confidence of 90%, and lift of 2.143. The retained rules had an average rule accuracy of approximately 50.25%. Rule-set reliability assessment showed completeness of 0.50 and consistency of 0.45, indicating constrained rule coverage and stability under the available incident data. Leave-one-source-out sensitivity analysis showed that three rules were stable, while one rule was partially stable. Structured expert review supported the interpretability and practical relevance of the retained rules for safety planning, Learning From Incidents (LFI), Job Safety Analysis (JSA), Permit-to-work (PTW) preparation, inspection checklists, and Design for Safety discussions.