Remote Motor Rehabilitation: EMG-IMU based Deep Learning Model Improves the Estimate of Wrist Kinematics

Abstract

Technology for motor rehabilitation faces challenges in uncontrolled settings, such as at home. In these real-world scenarios, robust signals like electromyographic (EMG) and inertial measurement unit (IMU) data are crucial for decoding continuous human actions. Classical modeling methods, such as linear, adaptive, or static filters, lack the capacity to capture complex relationships between surface EMG and kinematics, as well as generalizability across subjects. We propose a deep learning (DL) model for the continuous decoding of hand motion. Custom-made wearable devices acquired EMG-IMU data from 27 healthy subjects performing a wrist flexion/extension task. Two regression models were compared with a hybrid convolutional neural network and a gated recurrent unit. To address inter-subject variability, a leave-one-subject-out cross-validation approach was implemented. The DL model showed a mean R² increase of 0.18 compared to the polynomial regressor. Our approach enhances wrist kinematics decoding providing a generalized model based on data captured with wearable devices. The findings hold potential for innovative home-based telemedicine solutions in motor rehabilitation.