Compression-Based Feature Reduction for Upper Limb Motor Decoding from Peripheral Neural Signal

Objective: Neural prostheses generate an extensive volume of data during neural recording and decoding, presenting challenges in data processing and storage. A strategy to address this issue is compressing the data before transmission. Methods: This paper shows the possibility of efficiently compressing nerve data while achieving improved performance for an upper-limb neural prosthetic device. The new method, named Compression-Based Feature Reduction (CBFR), is divided into two stages: (1) applying a lightweight compression method to the original signal, and (2) reducing the number of features after feature extraction, thereby enhancing the classification accuracy in task-based experiments. Results: In an in vivo experiment involving patients with amputated hands, CBFR improved the classification performance of each finger movement up to 1.6% and 6.9% using Symmlet4 Discrete Wavelet Transform (DWT) and Walsh-Hadamard Transform (WHT), respectively, over the original signal. Both methods achieved up to 3x compression ratio from the original signal. Significane: The proposed data compression technique could help advance the field of neural prosthetics by addressing data transmission challenges and enhancing the performance of upper-limb neural prosthetic devices.