A generic neural network model to estimate populational neural activity for robust neural decoding.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Robust neural decoding is essential for advanced neural-machine interfaces.
• Predicting finger forces from neural activity is a key challenge.

Purpose of the Study:

• To develop a novel generic neural network model for continuous finger force prediction.
• To improve the reliability and intuitiveness of neural-machine interactions.

Main Methods:

• Implemented convolutional neural networks (CNNs) to map high-density electromyogram (HD-EMG) signals to motoneuron firing frequency.
• Extracted spatiotemporal features from EMG energy and frequency maps for improved learning efficiency.
• Developed a generic model trained on multi-participant data and used regression for real-time force prediction.

Main Results:

• The generic CNN model outperformed subject-specific neuron-decomposition and EMG-amplitude methods.
• Achieved higher correlation coefficients and lower prediction errors between measured and predicted forces.
• Demonstrated more stable force prediction performance over time.

Conclusions:

• The developed approach offers a generic and efficient continuous neural decoding method.
• Enables robust and real-time human-robot interactions.
• Advances the field of neural-machine interfacing.