Developing predictive models of construction fatality characteristics using machine learning

Construction fatalities have significant economic and emotional burdens to construction employees, families, and organizations. Understanding critical factors influencing construction fatalities and eventually developing predictive models to predict construction fatality characteristics are therefore important. Such activities, which are traditionally based on questionnaire and simple statistical analysis, can now be conducted using comprehensive datasets on construction fatality and advanced machine learning approaches. This study aims to develop predictive models of construction fatality characteristics, including nature of injury (NOI), part of body (POB), source of injury (SOI), and event or exposure (EOE) using machine learning approaches. 30 explanatory variables from 694 fatalities reported by the National Institute for Occupational Safety and Health from 1982 to 2014 are used to build the predictive models, with prediction accuracy of 56.6%, 54.0%, 76.5% and 84.9% for NOI, POB, SOI, EOE respectively. Specifically, the model has a prediction accuracy of 84.7% for construction fall fatalities. Important indicators for predicting SOI and EOE are largely the same, with the most important ones being the likelihood of fall, PFAS (personal fall arrest system, including its functionality and relevant training), workers' activity, onsite safety equipment and install safety protection. Similarly, important indicators for predicting NOI and POB include fall, PFAS, injury year, workers' activity, location and safety equipment. The results will offer useful guidance for construction organizations to establish relevant emergency response plans and first aid facilities and services that correspond to the most likely NOI, POB, SOI and EOE on construction sites.