CSAE enhances movement classification from SEMG signals

Classification of motion from surface electromyography (sEMG) signals remains a major challenge in myoelectric prosthesis control and is often limited by individual differences and complex sensor configurations. Blagoj Hristov, Zoran Hadzi-Velkov, and Katerina Hadzi-Velkova Saneva from the University of Ss. Cyril and Methodius, along with Gorjan Nadzinski, Vesna Ojleska Latkoska, and colleagues, developed a convolutional sparse autoencoder (CSAE) for robust gesture recognition with just two sEMG channels. We will present a new deep learning framework using . Their work is important because it avoids the need for manual feature engineering and demonstrates excellent accuracy over multiple subjects, reaching an F1 score of 94.3% on a six-class gesture set. Importantly, the transfer protocol dramatically improves performance for new users, and the gradual learning strategy allows easy expansion to more complex movement sets, paving the way for more affordable and adaptable prosthetics.

The research team achieved a multi-subject F1 score of 94.3% ±0.3% with a six-class gesture set, demonstrating the potential for robust and reliable prosthetic control with minimal sensor input.

This breakthrough technology relies on a convolutional sparse autoencoder (CSAE) that extracts temporal features directly from the raw sEMG signal, avoiding the need for complex manual feature selection. This study introduces a novel transfer learning protocol to overcome the challenges posed by intersubject variability, a common impediment to myoelectric systems.
Using minimal calibration data, performance for unseen subjects improved dramatically from 35.1% ±3.1% of the baseline to 92.3% ±0.9%, demonstrating significant advances in adaptability and personalized control. This multiple learning approach allows the system to quickly adapt to new users, reducing the time and effort required for individual adjustments.

Additionally, the system exhibits scalability through an incremental learning strategy, extending to a 10-class gesture set with an F1 score of 90.0% ±0.2% without retraining the complete model. By leveraging a convolutional architecture and LASSO regularization, CSAE learns a sparse and robust representation of neuromuscular activity, enabling efficient feature extraction and improved generalization.

This approach avoids the computational burden of recurrent neural networks and is suitable for real-time applications and resource-constrained hardware. This study establishes a scalable and efficient path toward affordable adaptive prosthetic systems that combine high accuracy with minimal computational and sensor requirements.

By learning directly from raw signals, this framework enables more intuitive and performant control, potentially changing the lives of individuals who rely on prosthetic limbs. This work opens new avenues for developing the next generation of prosthetics that are accessible and responsive to user needs.

Convolutional sparse autoencoder training and few-shot transfer learning for gesture decoding shows promising results

To address the limitations of myoelectric prosthesis control, scientists have developed a deep learning framework for accurate gesture recognition utilizing only two surface electromyography (sEMG) channels. The research team designed a convolutional sparse autoencoder (CSAE) that directly extracts temporal feature representations from raw sEMG signals, avoiding the need for manual feature engineering.

This innovative approach allows systems to learn directly from signals, potentially discovering more robust and expressive features to improve control. We used a 6-class gesture set in our experiments to evaluate the performance of our model and achieved a multi-subject F1 score of 94.3% ±0.3%. To reduce the effects of inter-subject variability, this study pioneered a few-shot transfer learning protocol and improved performance for unseen subjects from 35.1% ±3.1% of the baseline to 92.3% ±0.9% using minimal calibration data.

This transfer protocol allows rapid adaptation to new users and reduces the time and effort required for individual adjustments. The system further demonstrates scalability through an incremental learning strategy, extending to a 10-class gesture set while maintaining an F1 score of 90.0% ±0.2% without completely retraining the model.

The researchers used this technology to enable the system to learn new gestures without forgetting previously learned gestures, improving adaptability and long-term usability. The CSAE architecture was chosen specifically for its ability to preserve the temporal structure within the sEMG signal, unlike standard fully connected autoencoders that often require complex preprocessing. This method achieves high performance by processing raw time series data directly and avoiding the information loss associated with conversion to spectrogram or wavelet representations.

Deep learning and transfer learning improve accuracy and robustness to enhance multi-class myoelectric prosthesis control

Scientists achieved a multi-subject F1 score of 94.3% ±0.3% on a six-class gesture set using a deep learning framework for myoelectric prosthesis control. The research team developed a convolutional sparse autoencoder (CSAE) that extracts temporal features directly from raw surface electromyography (sEMG) signals, bypassing the need for manual feature engineering.

Experimental results reveal that CSAE effectively learns compact and meaningful representations of sEMG signals, which is important for real-time multi-motion prosthesis control. To address inter-subject variability, the team presented a transfer learning protocol that improved unseen subject performance from a baseline of 35.1% ±3.1% to 92.3% ±0.9% with minimal calibration data.

This adaptation process included a leave-one-subject strategy that split the data into training, validation, and test sets across eight subjects. Data standardization performed before model training ensured that there was no information leakage between sets and mimicked realistic prosthetic control conditions. The system demonstrated scalability through a stepwise strategy, achieving an F1 score of 90.0% ±0.2% on a 10-class gesture set without completely retraining the model.

The CSAE architecture consists of a symmetric encoder-decoder pair to process the sEMG input signal, represented as a matrix X∈RT×C. Here, T= 1000 time samples are collected over 250ms at a sampling rate of 4kHz using only two channels (C= 2). The encoder compresses the signal into a latent representation Z for information preservation, unsupervised learnability, and sparsity.

The researchers utilized strided convolution for learnable downsampling within the encoder to preserve temporal structure and improve computational efficiency. The objective function promotes sparsity by applying an L1 penalty to the activation of the bottleneck layer, promoting a disentangled and efficient representation of the input data. Measurements confirm that this model is able to learn robust feature representations from as few as two sEMG channels, paving the way for affordable adaptive prosthetic systems.

2-channel sEMG control uses deep learning for high precision and adaptability, providing intuitive prosthesis control

To address important limitations in myoelectric prosthesis control, scientists have developed a deep learning framework that uses only two surface electromyography (sEMG) channels for accurate gesture recognition. The proposed method uses a convolutional sparse autoencoder (CSAE) to directly extract temporal features from the raw signal, eliminating the need for manual feature engineering and simplifying the process.

The framework achieved a multi-subject F1 score of 94.3% on a six-class gesture set and demonstrated high accuracy with minimal sensor requirements. The researchers also introduced a transfer protocol that significantly improved performance for new users, increasing accuracy from 35.1% to 92.3% with limited calibration data.

The scalability of the system is confirmed by an incremental learning strategy, which successfully scales up to a 10-class gesture set with an F1 score of 90.0% without retraining the complete model. These findings challenge current industry trends relying on high-density sensor arrays and suggest a viable path toward more affordable and adaptable prosthetic systems.

The authors acknowledge that this study was initially conducted in healthy subjects and that further validation in a larger and more diverse population, including amputees, is required to confirm clinical relevance. Future research will focus on improving the system’s robustness against real-world challenges such as muscle fatigue and electrode shift, factors known to influence performance.

Furthermore, extending the functionality of the system to enable continuous proportional control is an important step toward creating more natural and intuitive neuroprosthetic devices. The public datasets used in the study will encourage further research and development in this area.