Deep Learning-Based Approaches for Decoding Motor Intent From Peripheral Nerve Signals

• DOI: 10.3389/fnins.2021.667907
• Publication Date: 2021-06-23T05:24:23Z
• Authors: Diu K. Luu; Anh T. Nguyen; Ming Jiang; Jian Xu; Markus W. Drealan; Jonathan Cheng; Edward W. Keefer; Qi Zhao; Zhi Yang
• Keywords: Sliding window protocol

Previous literature shows that deep learning is an effective tool to decode the motor intent from neural signals obtained different parts of nervous system. However, networks are often computationally complex and not feasible work in real-time. Here we investigate approaches' advantages disadvantages enhance learning-based decoding paradigm's efficiency inform its future implementation Our data recorded amputee's residual peripheral nerves. While primary analysis offline, nerve cut using a sliding window create “pseudo-online” dataset resembles conditions real-time paradigm. First, comprehensive collection feature extraction techniques applied reduce input dimensionality, which later helps substantially lower decoder's complexity, making it for translation Next, two strategies deploying models: one-step (1S) approach when big available two-step (2S) limited. This research predicts five individual finger movements four combinations fingers. The 1S recurrent network (RNN) concurrently predict all fingers' trajectories generally gives better prediction results than machine algorithms do same task. result reaffirms more advantageous classic methods handling large dataset. training on smaller set 2S approach, includes classification stage identify active fingers before predicting their trajectories, offer simpler while ensuring comparably good outcomes ones. In step, either or models achieve accuracy F1 score 0.99. Thanks regression both types comparable mean squared error (MSE) variance accounted (VAF) scores as those approach. study outlines trade-offs real-time, low-latency, high decoder clinical applications.