Deep Learning-Based Robust Neural-Machine Interface for Dexterous Control of Robotic Hand

Neuromuscular injuries can impair hand function and impact the quality of life. To restore hand dexterity, numerous assistive devices have been developed. However, the lack of a robust neural-machine interface may limit functionality of these devices. Accordingly, a robust neural decoding approach was developed that can continuously decode the intended finger motor output. High-density electromyogram (HD-EMG) signals were obtained from the extrinsic finger flexor and extensor muscles. Convolutional neural networks were implemented to learn the mapping from HD-EMG features to finger-specific population neuron firing frequency, which was then used to control a prosthetic hand in real-time. In comparison with the HD-EMG amplitude approach, the network-based decoder predicted finger forces and angles with lower prediction errors. The network-based decoder also demonstrated better isolation with minimal predicted output in the unintended fingers. The outcomes offer a novel neural-machine interface technique that allows intuitive control of assistive robotic hands in a dexterous manner.Master of Scienc