In construction sites, the diversity of project types and work activities leads to the recurrent occurrence of various forms of occupational accidents each year. To effectively prevent such incidents, it is essential to analyze past accident cases, pr...
In construction sites, the diversity of project types and work activities leads to the recurrent occurrence of various forms of occupational accidents each year. To effectively prevent such incidents, it is essential to analyze past accident cases, predict potential accident types through data-driven approaches, and establish preventive measures based on quantitative indicators. Accordingly, this study aims to develop an artificial intelligence (AI)-based prediction model capable of forecasting accident types in construction sites, and to utilize the results as a quantitative decision-making tool within the framework of risk assessment (RA).
The dataset used in this study consisted of 320 accident cases collected over six years (2019–2024) and 48 new cases from 2025. Each accident record was systematically structured using standardized occupational, task, and accident-type codes. To address the severe class imbalance inherent in the data, a hybrid resampling strategy combining undersampling and SMOTENC was applied. The predictive model was developed using a Soft Voting Ensemble framework integrating Random Forest (RF), XGBoost (XGB), and LightGBM (LGBM) classifiers. To enhance the reliability of predicted probabilities, Platt Scaling with a sigmoid function was employed for probability calibration. Furthermore, a hierarchical classification approach reflecting the 19-category accident taxonomy was implemented, along with threshold tuning for class-wise probability adjustment, to improve both model performance and practical applicability.
Through a stepwise performance enhancement process, the proposed ensemble and probability-calibrated model demonstrated improved stability and reliability compared to single-model baselines. The final model achieved approximately 80% Top-5 accuracy, with notable improvements in Mean Reciprocal Rank (MRR), Brier Score, and Expected Calibration Error (ECE), indicating both enhanced predictive precision and probabilistic consistency. When applied to 48 new accident cases from 2025 for external validation, the model achieved high hit rates for major accident categories, thereby confirming its potential applicability to real-world construction safety management.
This study represents the first attempt to quantitatively structure and apply real accident data from a construction company to an AI-based accident-type prediction model, highlighting the potential to establish a data-driven decision-making framework for construction safety management. In particular, by utilizing the model’s Top-5 predicted probabilities as the “frequency” component within the risk assessment process, this research demonstrates the feasibility of implementing an AI-driven Dynamic Risk Assessment (DRA) capable of quantifying task-level risks and prioritizing preventive measures. Future research should integrate diverse accident datasets across multiple construction projects and combine deep learning algorithms with real-time data acquisition systems to further enhance the model’s generalizability and on-site applicability.
Key words: construction accident prediction, artificial intelligence (AI), AI enhancement techniques (data imbalance, ensemble learning, probability calibration, hierarchical classification, threshold tuning), risk assessment