This study included motion tracking data collected from 15 healthy participants (eight females and seven males; mean age, 25±5 years) with no musculoskeletal or gait disorder. The mean height and weight were 170.6±8.9 cm and 65.8±12.1 kg, respectively. All participants confirmed their right-hand dominance.

Three markerless depth sensors (Azure Kinect, Microsoft) for motion tracking were placed on the flat ground. All depth sensors had zero-angle tilting and were mounted on individual tripods. The treadmill was placed at the center of the three cameras. The handrail of the treadmill was removed to prevent camera occlusion (Fig. 1). The distance (2 m) and height (1.5 m) of the cameras from the treadmill were fixed to ensure that the entire body was within the operating range of each sensor as suggested by the manufacturer [12]. The frontal depth sensor pointed directly toward the treadmill front for the full-frontal plane view. The other two sensors pointed to the left and right sides of the treadmill (toward left and right arms) orthogonal to the frontal camera line of sight (Fig. 1).

Motion during the experiment was captured at 30 Hz, and data acquisition proceeded using markerless motion capture software (iPi Motion Capture, LLC) (Fig. 2A). Inverse dynamics were computed from the kinematic measurement values (ROM, position, velocity, and acceleration of elbow and shoulder) of the depth sensor data. The elbow was considered to be set at 2 degrees of freedom (DOF), its flexion/extension and pronation/supination angles were measured, and moments were calculated for sagittal and axial planes, respectively. The shoulder joint was treated as a 3-DOF joint; its flexion/extension, abduction/adduction, and external rotation/internal rotation angles were measured; and moments were calculated for sagittal, coronal, and axial planes, respectively. When the three depth sensors started body tracking, their respective PCs independently sent data to the main computer that collected data for synchronization.

Subjects were instructed before the start of the experiment to warm up by walking at a constant speed controlled by the treadmill (walking, 1.11 m/sec) for 60 seconds to become familiar with the apparatus. The constant speed for walking was selected based on a reference gait dataset for a 20- to 39-year-old age group [13]. The reference gait patterns of the speed (walking, 30 seconds per trial) were recorded by depth-censored cameras.

The specific X-Y-Z locations obtained from the motion capture session were used as input to the musculoskeletal model in the AnyBody Modeling System software (ver. 7.3.0, AnyBody Technology). A global axis specified the environmental factor of gravity (−9.8 m/sec² in the superior–inferior direction). The upper body portion of the full-body model from the AnyBody Managed Model Repository (ver. 2.3.0) was used. The model consists of multiple bones comprising the pelvis, trunk, neck, and arms. The arm consists of several joints including three spherical joints (sternoclavicular, acromioclavicular, and glenohumeral), three revolute joints (elbow and two in the wrist), and pronation/supination of the forearm. The model also includes lumbar, thoracic, and cervical spine regions and contains multiple muscle fascicles of upper body muscles (Fig. 2B). The generic model was scaled to each subject using anthropomorphic data. The muscle groups were sectioned into force vectors connected at two points in the body to represent the origin and insertion points of the muscle. Moreover, model parameter identification and inverse dynamic analysis were conducted following a previously validated study [14]. Inverse dynamics is a concept used primarily in biomechanics, robotics, and control engineering to analyze and understand the forces and torques involved in the motion of a system. It refers to the process of working backward from the desired motion or trajectory of a system to determine the required forces and torques at each joint or point of interaction. In contrast to forward dynamics, in which forces and torques are used as inputs to calculate the motion of the system, inverse dynamics starts with the motion (position, velocity, and acceleration) and calculates the forces and torques required to achieve that motion.

Data analysis was completed using MATLAB. The data were segmented into trials, where segmented gait patterns were normalized to the gait cycle. Seven gait cycles in the middle of each 30 seconds recording were selected and averaged for data analysis. A gait cycle was defined as the movement from touchdown of the heel of the reference foot on the ground to the endpoint of the terminal swing of the same foot (Fig. 3B).

All postacquisition data processing was performed in Python 3.7.4. After averaging the normalized kinematic gait patterns of the 10 sampled steps, the ROM was computed in terms of maximum and minimum joint angles for each trial. The mean joint moment of each time frame in each joint was calculated in Newton meters (Nm).

Statistical analysis was performed under the supervision of a biostatistician. The Kolmogorov-Smirnov test was used for normality distribution for all datasets. All descriptive and quantitative analyses were conducted using SPSS ver. 22.0 (IBM SPSS Corp.).

The mean ROM of the dominant elbow during a gait cycle was 29.7°±10.2° and 14.2°±3.2° in flexion–extension and pronation–supination, respectively (Fig. 3A). The mean ROM of the dominant of the dominant elbow during a gait was 56.4°±12.7°, 25.6°±5.2°, and 19.8°±4.6° in flexion–extension, rotation, and abduction–adduction, respectively.

The mean joint moment (in Nm) of the dominant elbow during a gait cycle was 33.5±9.2 and 1.2±0.4 Nm in flexion–extension and pronation–supination, respectively (Fig. 3A). The mean joint moment of the dominant shoulder during a gait cycle was 91.3±22.6, 11.0±4.3, and 25.5±12.0 Nm in flexion–extension, rotation, and abduction–adduction, respectively.