Prediction of Center-of-Mass Kinematics of Sensopro Exercises with Neural Network Models

Monitoring center-of-mass is crucial for assessing postural control, but field measurements are often impractical or cost-prohibitive. This study investigates the feasibility of predicting center-of-mass kinematics from the motion of an unstable base—the Sensopro Luna—using deep learning, eliminating the need for wearable sensors. We conducted a cross-sectional study in which 64 participants were recorded performing three coordination exercises (Single-Leg Stance, Stepping, and Waves). Marker-based motion capture and auxiliary inertial sensors were used to record reference and tape kinematics. The model inputs consisted of IMU- and motion-capture-derived tape segment orientations, IMU accelerations and angular velocities, and algorithmic estimates of the lowest tape positions. Nine axis-specific exercise models were developed using a hybrid Encoder–LSTM–Decoder architecture and compared against linear regression baselines. Our results indicate that the deep learning models successfully predicted horizontal center-of-mass displacements (DNN Mean Absolute Errors of 16.1–23.7 mm for X-axis and 4.4–31.3 mm for Y-axis) and exhibited descriptively lower errors than linear models in mean absolute error and signal morphology. However, vertical predictions were less reliable, likely due to the physical constraints inherent to the kinematics of the unstable base. Error analysis revealed that prediction accuracy was highest within common postural ranges, but decreased for extreme displacements. These findings provide a proof-of-concept for wearable-free postural monitoring, particularly for movement along the mediolateral and sagittal axes. Such a system could facilitate automated, cost-effective postural feedback and performance tracking in rehabilitation and fitness environments, supporting autonomous coordination training without the practical constraints of traditional measurement systems.