What the paper contributes
This study tests whether structured occupational accident records can support prediction of factors associated with fatal falls. Seven machine-learning models were compared using 3,321 accident scenarios, and Random Forest was selected as the principal interpretable classifier.
Method
The dataset combined date, activity, region, accident summaries, direct and root causes, and contributing factors. Models were evaluated across agent, management, site-condition and individual-characteristic factor groups.
Main finding
The work established the feasibility of machine learning for construction-safety management while showing that predictive performance changes across factor groups. Feature importance provided an interpretable view of what shaped the predictions.
Practical value
The model is not a substitute for investigation. Its value is in screening large incident datasets, increasing awareness of potential hazards and helping prioritise prevention questions.