Predicting serious injury and fatality exposure using machine learning in construction projects

Safety academics and practitioners in construction typically use safety prediction models that employ information associated with past incidents to predict the likelihood of future injury or fatality on site. However, most prevailing models utilize only information related to failure (i.e., incident), so they cannot distinguish effectively between success and failure without well-informed comparison. Furthermore, recordable incidents on construction sites are extremely rare, which results in data that are too sparse to make predictions with high statistical power. This paper empirically reviews different approaches to safety to increase the understanding of conditions associated with safety success and failure. Empirical data about business-, project-, and crew-related factors were collected to predict serious injury and fatality (SIF) exposure conditions. A variety of modeling techniques were tested in a machine learning pipeline to identify the most accurate and stable predictive models. Results showed that the multilayer perceptron (MLP) approach best distinguished SIF exposure conditions from safety success conditions using nonlinear decision boundaries. The most influential factors in the models included the crew experience working together, supervisor experience with the crew, total number of workers under the supervisor's purview, and the maturity of leadership development programs for frontline supervisors. This study showed that data sets with both success and failure information yield more reliable and meaningful predictions than data sets with failure alone. Such an approach to safety data collection, analysis, and prediction could be used by future researchers to generate new insights into the causes of serious incidents and the relationships among causal factors.