Assessment of Synergy-Assisted EMG-Driven NMSK Model for Upper Limb Muscle Activation Prediction in Cross-Country Sit-Skiing Double Poling

Xue Chen¹Zhongxue Yuan²Xianzhi Gao¹Yanxin Zhang²Chenglin Liu^3*Bo Huo^3*

• ¹Beijing Institute of Technology, Beijing, Beijing Municipality, China
• ²The University of Auckland, Auckland, Auckland, New Zealand
• ³Capital Institute of Physical Education and Sports, Beijing, China

Cross-country sit-skiers are often individuals with spinal cord injuries, cerebral palsy, or lower limb disabilities, relying heavily on upper limb strength to generate propulsion during skiing. However, frequent shoulder joint movements significantly increase the incidence of shoulder joint disorders. Therefore, quantifying muscle forces during movement is crucial for understanding upper limb force generation patterns. Currently, electromyography (EMG)-driven neuromusculoskeletal (NMSK) models are the predominant method for calculating muscle forces and joint moments. However, this approach heavily depends on the quality and quantity of EMG data. Surface electrodes are typically used to collect activity data from superficial muscles, but during dynamic movements, factors such as skin stretching, sweating, or friction may cause electrode detachment or poor contact, leading to EMG signal acquisition failures or data loss. In this study, we propose a synergy-assisted EMGdriven NMSK model to predict the activation patterns of missing muscles for cross-country sit-skiing double poling. This method is based on individualized EMG-driven NMSK models constructed for each participant, incorporating data from 10 muscles. By utilizing the activation data of 9 known muscles, the model predicts the activation of one missing muscle through synergy analysis. For synergy method selection, we systematically compared four approaches: Non-negative Matrix Factorization (NMF), Principal Component Analysis (PCA), Independent Component Analysis (ICA), and Factor Analysis (FA). The results demonstrated NMF's superior performance at 5 synergies, accurately predicting any missing muscle activation among 10 muscles ( r = 0.79 ± 0.25 vs. 0.14 ± 0.60-0.45 ± 0.63 for alternatives, p < 0.05), with lower errors (RMSE: 0.21 ± 0.11, p < 0.05 vs. ICA/FA, p < 0.1 vs. PCA; MAE: 0.17 ± 0.09, all p < 0.05). This finding validates the effectiveness of the proposed method in predicting upper limb muscle activation during coupled shoulder and elbow joint movements.