A SYSTEMATIC REVIEW OF ARTIFICIAL INTELLIGENCE BASED PREDICTIVE SAFETY MODELS FOR REDUCING WORKPLACE INJURIES IN MANUFACTURING AND CONSTRUCTION

Abstract

This study presented a systematic review of artificial intelligence–based predictive safety models aimed at reducing workplace injuries in manufacturing and construction, with emphasis on quantitative comparability across outcomes, data modalities, validation designs, and performance metrics. A total of 312 observational units and respondent-linked records from manufacturing and construction contexts were synthesized to evaluate injury occurrence, high-severity injury outcomes, and leading-indicator–based risk prediction. Manufacturing accounted for 51.9% of the analyzed records, while construction represented 48.1%. Descriptive results showed moderate-to-high levels of perceived AI usefulness (mean = 3.92, SD = 0.64) and leading-indicator maturity (mean = 3.74, SD = 0.69), with construction exhibiting higher median near-miss activity (median = 2 events per unit window) than manufacturing (median = 1). Logistic regression analyses indicated that data quality readiness was significantly associated with reduced injury occurrence (odds ratio = 0.78, p = 0.002) and reduced high-severity injury occurrence (odds ratio = 0.73, p = 0.006). Safety culture also demonstrated a protective association with injury occurrence (odds ratio = 0.82, p = 0.013). Sector-stratified analyses showed stronger readiness effects in construction (odds ratio = 0.72, p = 0.001) than in manufacturing (odds ratio = 0.83, p = 0.041). Leading-indicator maturity was associated with lower general injury odds (odds ratio = 0.85, p = 0.028) but did not reach significance for high-severity injuries. Validation design and metric selection were found to substantially influence reported performance, with temporal and site-held-out testing yielding more conservative and credible estimates than random splits. Overall, the findings underscored that predictive safety effectiveness depended primarily on data readiness, measurement quality, and validation rigor rather than algorithm complexity alone.