Overloading and overusing of human body joints and their active–passive structures during occupational activities play a crucial role in the etiology of musculoskeletal disorders (Thiese et al., 2014, Van Nieuwenhuyse et al., 2004). To assess and manage risk of musculoskeletal injuries during these activities both ergonomics risk assessment tools (e.g., RULA, REBA, OWAS, and NIOSH) (Hignett and McAtamney, 2000, McAtamney and Nigel Corlett, 1993, Waters et al., 1994, 1993) and biomechanical musculoskeletal models (Arjmand et al., 2013, Calder and Potvin, 2012, Damsgaard et al., 2006, Delp et al., 2007, Khoo et al., 1995, Subramaniyam et al., 2015) are employed. Moreover, musculoskeletal models, that predict joint loads and muscle forces, have important applications in surgical decision making, patients’ pre/post-surgery assessments, non-surgical intervention assessments, orthoses assessments, and design of rehabilitation programs (Ebrahimkhani et al., 2021, Ignasiak, 2020, Reeves and Cholewicki, 2003, Smith et al., 2021). An essential input to both ergonomics risk assessment tools and musculoskeletal models is the three-dimensional (3D) human body posture during the activity under consideration.

Body posture is either measured experimentally or predicted mathematically. Measurement approaches are usually via medical imaging (e.g., MR and fluoroscopy) (Eskandari et al., 2017, Zanjani-Pour et al., 2017), skin-based (e.g., optical, inertial, or magnetic) motion capture techniques (Arjmand et al., 2009, Cholewicki and McGill, 1996, Faber et al., 2016, Hajibozorgi and Arjmand, 2016, Hsu et al., 2008), and marker-less depth/RGB cameras (Asadi and Arjmand, 2020, Hong and Kim, 2018). While these measurement approaches are costly, time-consuming, and/or limited to research laboratories, their accuracy can be compromised when body joints are occluded, for instance, by an object in the real workstation. Therefore, posture-prediction approaches, i.e., link segment models based on an inverse-kinematics algorithm, have been developed to predict posture (link segment orientations) for a given hand and foot position (Dysart and Woldstad, 1996, Marler et al., 2011, Woldstad, 1997). As such linkage models are kinematically redundant, i.e., there exist many different possible postures (link orientations) for a given hand and foot position, an optimization approach should be used to predict a mathematically optimal posture that could indeed be very different from the actual posture.

As alternatives, artificial intelligence-based posture-prediction approaches have emerged (Aghazadeh et al., 2020, Gholipour and Arjmand, 2016, Li et al., 2021, Mohseni et al., 2022, Mohseni et al., 2023, Mousavi et al., 2020, Perez and Nussbaum, 2008, Zhang et al., 2010). In these posture-prediction approaches, posture data from several individuals are first measured via a skin-based motion capture system during various physical/occupational activities, and a machine-learning tool such as an artificial neural network (ANN) is subsequently trained on these measured data for the generalization purposes. Our most recent ANN is trained based on a comprehensive experimental database of 41 plug-in-gait full-body skin markers collected from twenty subjects, each performing 204 static symmetric/asymmetric one/two-handed activities in upright standing and flexed postures (Mohseni et al., 2023). This easy-to-use ANN is capable of predicting the 3D coordinates of the skin markers when the individual statically holds a hand-load. It can accurately predict body joint coordinates during various activities with a root-mean-square-error of ∼ 2.5 cm (averaged for all the skin marker coordinates); a considerable improvement when compared to our (Aghazadeh et al., 2020, Gholipour and Arjmand, 2016, Mohseni et al., 2022, Mohseni et al., 2023, Mousavi et al., 2020) and others’ (Li et al., 2021, Perez and Nussbaum, 2008) previous posture-prediction ANNs.

This and other posture-prediction ANNs are, however, limited to predicting body posture in static load-handling activities. To train an ANN that can predict body posture during dynamic lifting activities, in which a weight is lifted from an origin to a destination, one needs to measure body posture of several individuals performing lots of dynamic lifting activities with different hand-load origins and destinations. Such measurements and the associated experimental setup are, however, time-consuming and strenuous. Alternatively, to predict posture during dynamic activities, one may use the existing posture-prediction ANNs that are developed based on static load-handling postures, as any dynamic movement can be broken down into several sequential static frames that can, in turn, be predicted by our ANN. For this, the 3D hand-load coordinates during the dynamic task under consideration are first input into the ANN as functions of time and body posture is subsequently predicted in each time frame so to generate a dynamic body movement. Robustness of the existing statically-developed ANNs in predicting body posture during dynamic lifting activities remains to be investigated. Moreover, effects of the ANN posture-prediction errors on the joint load estimations of musculoskeletal models during dynamic lifting activities have not been evaluated either.

The present study, therefore, aims to investigate the capability of our ANN (Mohseni et al., 2023) to predict 3D body posture during dynamic lifting activities. Seven individuals each performed twenty-five symmetric/asymmetric one- and two-handed dynamic lifting activities while their full-body posture (i.e., coordinates of skin markers) was measured via a ten-camera VICON motion capture system. To predict posture in these dynamic tasks, the time-dependent hand-load position was input into the ANN, and its capability to predict posture was assessed by comparing the predicted and measured postures. Moreover, effects of the ANN posture-prediction errors on spinal loads during these dynamic tasks, as required for occupational risk assessments, were investigated using subject-specific musculoskeletal models driven by either measured or predicted postures.