Breakthrough in sEMG-Based Knee Angle Prediction
Researchers have developed a novel physiology-inspired bidirectional state space network, known as SynSSM-Net, that significantly improves continuous prediction of knee joint angles from surface electromyography signals. The work, published in Engineering Applications of Artificial Intelligence, addresses longstanding challenges in rehabilitation robotics and assistive technologies by incorporating muscle synergy principles, adaptive temporal alignment, and multi-scale feature extraction.
The study is led by Bin Feng, Liuyi Ling, Huashun Li, Liao Fang, and Zhipeng Yu. Their approach outperforms traditional data-driven models by embedding physiological constraints directly into the neural architecture. This results in more accurate, subject-specific predictions that could enhance real-time control of exoskeletons and prosthetic devices.
Understanding Surface Electromyography and Its Role in Joint Prediction
Surface electromyography, or sEMG, records electrical activity produced by skeletal muscles through electrodes placed on the skin. These signals precede actual muscle contraction and joint movement by tens of milliseconds, offering a valuable anticipatory input for motion prediction systems. In rehabilitation settings, accurate sEMG interpretation allows devices to anticipate user intent and provide timely assistance.
Continuous knee joint angle prediction from sEMG remains difficult due to nonlinear relationships, inter-subject variability, electrode placement sensitivity, fatigue effects, and cross-talk between muscles. Traditional biomechanical models like Hill-type muscle models require extensive individual calibration, while purely data-driven methods often overlook underlying physiological structure.
The SynSSM-Net Architecture: Integrating Physiology and Deep Learning
SynSSM-Net combines several innovative components to overcome these limitations. A softmax-constrained Learnable Delay Layer provides regularized temporal alignment across sEMG channels, compensating for electromechanical delays that vary by muscle and condition. This adaptive mechanism replaces fixed delays used in earlier approaches.
The core Synergy-SSM module couples muscle synergy decomposition with bidirectional state space modeling. It projects multi-channel sEMG into a low-dimensional, non-negative matrix factorization-warm-started synergy representation inspired by muscle synergy theory, where a small number of coordinated activation patterns explain most variance in muscle activity during locomotion.
Additional modules include Temporal–Statistical Selective Kernel (TF-SK) for adaptive temporal-statistical fusion and channel-temporal attention mechanisms. A physiology-constrained loss function incorporates NMF-based reconstruction consistency and Hoyer sparsity regularization to enforce physiologically plausible synergy activations.
Performance on Self-Collected and Public Datasets
Evaluations used a self-collected dataset from six healthy adults and the public Exoskeleton and Prosthetic Intelligent Controls (EPIC) dataset. Under subject-specific offline protocols with a 40-millisecond prediction horizon, SynSSM-Net achieved a coefficient of determination of 0.915 and root mean square error of 4.31 degrees on the self-collected data. On the main EPIC subset, it reached 0.920 and 5.36 degrees, outperforming eleven fixed-configuration baselines.
Ablation studies confirmed the contribution of each component, with the synergy bottleneck and bidirectional state space transitions providing the largest gains. Multi-horizon analysis demonstrated robustness across different prediction windows, while a pathological-cohort evaluation on University of California Irvine data yielded strong results with a coefficient of determination of 0.872.
Photo by Ortopediatri Çocuk Ortopedi Akademisi on Unsplash
Implications for Rehabilitation Robotics and Prosthetics
Improved continuous knee angle prediction enables more responsive and natural control of lower-limb assistive devices. Rehabilitation robots can use these anticipatory signals to synchronize assistance with user intent, reducing compensatory movements and improving training outcomes for individuals recovering from stroke, spinal cord injury, or joint replacement surgery.
In prosthetic applications, the network's structured representation learning supports better generalization across gait phases and activities. By respecting muscle synergy principles, the model produces predictions that align more closely with natural neuromuscular coordination, potentially lowering metabolic cost and enhancing user comfort during daily activities.
Broader Context in Biomedical Signal Processing
This work builds on decades of research in sEMG-based motion intention recognition. Earlier efforts relied on hand-crafted features and classical machine learning, while recent deep learning advances have focused on recurrent, convolutional, and transformer architectures. SynSSM-Net distinguishes itself by explicitly embedding physiological knowledge rather than treating sEMG purely as a generic time series.
The approach also advances state space modeling techniques in biosignal applications. Bidirectional processing captures both forward and backward temporal dependencies, complementing the synergy-informed dimensionality reduction that makes computation more efficient and interpretable.
Future Directions and Clinical Translation
While offline subject-specific results are promising, further validation in online, multi-subject, and real-world deployment settings is needed. Integration with inertial measurement units or other modalities could further improve robustness. Researchers highlight the potential for extending the framework to additional lower-limb joints and pathological populations.
The physiology-constrained design also opens avenues for interpretability studies, allowing clinicians to examine how synergy activations relate to specific movement impairments. Such insights could inform personalized rehabilitation protocols.
Funding and Institutional Support
The research received support from the Research and Development Fund of the Institute of Environmental Friendly Materials and Occupational Health at Anhui University of Science and Technology, along with provincial academic support programs in Anhui Province, China. These resources enabled the development and testing of the SynSSM-Net framework.
Photo by Navy Medicine on Unsplash
Accessing the Original Publication
The full paper, titled "A physiology-inspired bidirectional state space network for continuous knee joint angle prediction from surface electromyography," appears in Engineering Applications of Artificial Intelligence, Volume 181, Part 3, October 2026. Readers can access the abstract and details at the ScienceDirect page. The authors are Bin Feng, Liuyi Ling, Huashun Li, Liao Fang, and Zhipeng Yu.
