A Prototype Predictive Risk Model for Transferable Safety Analytics in Construction SMEs and LMICs

Predictive safety analytics has gained increasing attention in construction research; however, its practical uptake remains limited in small and medium-sized enterprises (SMEs) and construction sectors in low- and middle-income countries (LMICs), where incident reporting is fragmented, and structured safety data are scarce. This study addresses this challenge by developing a transferable prototype for activity-level risk classification to support proactive safety decision-making in data-constrained environments. Publicly available construction incident data reported under Health and Safety Executive’s Reporting of Injuries, Diseases and Dangerous Occurrences Regulations (RIDDOR) framework for Great Britain are used as a methodological testbed to design and evaluate the model architecture, without assuming direct transferability of incident patterns across contexts. The approach integrates systematic feature engineering and supervised machine-learning classification to categorise construction activities into risk levels using incident-level predictors. Several machine learning and deep learning models were evaluated, including Random Forest, Support Vector Machine, XGBoost, and Artificial Neural Networks. Results show that ensemble and neural models achieve moderate classification accuracy with strong discriminative performance, with AUC values above 0.85. These findings indicate that activity-level safety risk classification can support prioritisation of higher-risk activities but is better suited for relative risk ranking than precise risk prediction. Overall, the study contributes a transferable framework for safety analytics that can be refined and validated using local datasets.