Critical review on data-driven approaches for learning from accidents: Comparative analysis and future research

Data-driven intelligent technologies are promoting a disruptive digital transformation of human society. Industrial accident prevention is also amid this change. Although many emerging technologies, such as machine learning (ML), are extensively employed in workplace safety, these approaches need to fit the intended safety purpose of accident analysis, risk assessment, adverse outcome prediction, or anomaly detection. Hence, examining the "real-world" need for accident prevention and the advantages of emerging data-driven methodologies to better integrate them is necessary. This study provides a systematic review to clarify the current research status, existing problems, and future insights into these evolving technologies in accident prevention. We present notable gaps and barriers in data-driven accident prevention by analyzing 194 published studies from four perspectives: Paradigm, Model, Data Source, and Purpose. The results demonstrate (1) lack of a systematic framework to guide the application of Big Data (BD) in the field of safety; (2) few prior studies have considered model interpretability; (3) more proactive data needs to be incorporated into accident analysis; (4) safety-related data and domain knowledge need to be further integrated; (5) some recent data-driven techniques are unexplored in safety science. Further, the future research opportunities are discussed based on these findings. Such review may help clarify the mapping of data-driven tasks to safety goals to accelerate the uptake of data-driven technologies in safety or accident analysis research.