A Deep Learning based Model for Decoding Motion Intent of Traumatic Brain Injured Patients' using HD-sEMG Recordings

Abstract

Traumatic brain injury (TBI) post-stroke survivors mostly face the challenges of performing key daily life activities. Recent studies have attempted to develop intelligent rehabilitation robotic systems to engage patients in active motor training for rapid functional recovery. The systems are controlled by electromyography (EMG) signals based on a pattern recognition (PR) scheme to decode patients' motor intent that serves as their control input. However, most existing PR-based rehabilitation systems for TBI patients were developed on traditional approaches that use handcrafted features. As such, they only provide limited neural information for TBI patients' motor intent decoding. In addressing this problem, this study proposes a deep learning-based fully connected convolutional neural network (FC-CNN) model that automatically extracts rich set of motor information from high-density surface EMG (HD-sEMG) recordings of TBI post-stroke patients. Experimental results show that the proposed FC-CNN model could achieve consistently accurate decoding outcomes (above 96%) across different window sizes (100 ms, 150ms, 200ms, 250ms, and 300ms) and subjects. Also, an accuracy in the range of 90.08%~ 98.92% was recorded across subjects using a single experimental trial. The proposed model may facilitate the development of intuitively active rehabilitation robots for TBI post-stroke patients' functional recovery.