Spiking neural networks cut energy 200x for muscle fatigue detection

Surface electromyography (sEMG) is widely used to detect muscle fatigue in sports, rehabilitation, and wearable health monitoring. Accurate fatigue detection helps prevent injuries and optimize performance, but existing deep learning models are too computationally expensive and data-hungry for real-time wearable use. Researchers from the National University of Singapore (NUS) have now proposed a spiking neural network (SNN) framework that slashes energy consumption by up to 201.77x compared to conventional neural networks while matching or exceeding accuracy.

The team introduced a quantization-compatible training scheme called SDH, which incorporates multiple regularization terms to improve robustness under noisy conditions. They evaluated their quantized SNNs on two public sEMG datasets against a broad set of baselines and under seven different noise conditions, including physically motivated perturbations. Despite extreme energy efficiency, the SNNs remained more stable under diverse noise than standard deep learning models, making them ideal for low-power wearable devices that must operate reliably in real-world environments.

This breakthrough could unlock continuous muscle fatigue monitoring in smart sportswear, prosthetic control, and rehabilitation systems where battery life and real-time processing are critical. The paper, currently on arXiv, represents a significant step toward practical neuromorphic computing in healthcare and human-machine interfaces. The authors have open-sourced their code and results, inviting further development in energy-efficient AI for biomedical applications.