Muscle synergy-driven deep learning for predicting multiphase knee joint loading in ACL-reconstructed patients.

BACKGROUND

Accurate estimation of knee contact force (KCF) is valuable for monitoring joint loading during rehabilitation after anterior cruciate ligament reconstruction (ACL-R). Conventional musculoskeletal simulations, such as OpenSim, provide an effective approach for estimating KCF but typically require complex modeling workflows and substantial computational resources. This study aimed to evaluate the feasibility of integrating muscle synergy-derived neuromuscular representations within a neural network (SSA-CNN-xLSTM) framework to reproduce OpenSim simulation-derived KCF across different stages of ACL-R rehabilitation.

METHODS

Thirty-three individuals who underwent unilateral ACL-R were evaluated at 3, 6, and 12months postoperatively during level walking. Surface electromyography and motion capture data were collected, and muscle synergies were extracted using non-negative matrix factorization (NNMF). Two neuromuscular input representations, muscle activation signals and muscle synergy matrices, were used to train SSA-CNN-xLSTM surrogate models for predicting OpenSim simulation-derived KCF.

RESULTS

Both neuromuscular representations enabled accurate reproduction of OpenSim simulation-derived KCF across all rehabilitation stages. The proposed framework achieved high predictive agreement with simulation-derived reference outputs, with R² values exceeding 0.90 for all models. Under the corresponding SSA-optimized model configurations, muscle synergy-based models demonstrated comparable predictive performance and numerically higher agreement than muscle activation-based models at several rehabilitation stages. These findings demonstrate that low-dimensional neuromuscular representations can effectively support surrogate modeling of simulation-derived KCF throughout postoperative recovery.

CONCLUSION

The proposed SSA-CNN-xLSTM framework demonstrates the feasibility of reproducing OpenSim simulation-derived KCF using non-invasive neuromuscular measurements across multiple stages of ACL-R rehabilitation. The longitudinal study design provides evidence that the framework remains applicable under progressively changing neuromuscular conditions during recovery. These findings support the potential of muscle synergy-informed surrogate modeling as a complementary approach for biomechanical assessment, while further validation under more demanding functional tasks and quantitative evaluation of computational efficiency are warranted.