An automatic procedure to classify biomechanical risk classes associated to lifting activities according to the Revised NIOSH Lifting Equation

Introduction: In the field of occupational medicine, physical ergonomics studies, aimed at preventing work-related illnesses, have gained importance in recent years since both quantitative/semi-quantitative observational studies and instrumental methods have reported promising results. Among the instrumental methods, wearable devices capable to acquire signals such as acceleration and angular velocity, electromyography (EMG) [1] have potentially proved useful to help assessing - both automatically and more accurately - the work-related risk, which can lead to biomechanical overloading and the development of work-related musculoskeletal disorders (WRMD), especially in the case of workers involved in activities such as material handling and lifting tasks. Methods: In this context, a powerful method to monitor workers biomechanical risk is a National Institute for Occupational Safety and Health (NIOSH) quantitative formula (RNLE), which considers factors such as intensity, duration and frequency (together with other geometrical characteristics) of a lifting task [2]. The study was carried out enrolling 13 healthy subjects with no signs of musculoskeletal disorder wearing an inertial sensor (Opal, APDM Inc.) on the chest. Each subject performed a session based on two trials. Each trial consisted in 20 consecutive lifting of a plastic container. The first trial was performed in a condition with LI < 1 which corresponds to the No-risk class while the second trial was performed in a condition with LI > 1 which correspond to the Risk class. LI of 0.5 and 1.3 were derived from the RNLE by variously combining height, frequency and weight of lifting tasks. The collected acceleration and angular velocity signals along the three axes were filtered and then segmented to extract the region of interest (ROI) corresponding to the lifting by means of the Savitzky-Golay filter. For each ROI several features in the time and frequency domains were extracted and feed to machine learning algorithms to evaluate the discriminative power of these features to classify risk classes according to the RNLE. Results: Table 1 shows the results of the machine learning (ML) analysis by means of some evaluation metrics using different ML algorithms. The ML analysis was carried out on a dataset composed of 520 instances (13 subject x 40 lifting), 114 features (19 features x 6 axes) and two classes (No-risk, risk).DiscussionThe goal of this research was to develop an automatic procedure - base