Multi-Gesture Decoding using Hybrid EMG-IMU for Rehabilitation Applications

Abstract

Prosthetic rehabilitation of the upper limb through powered prosthetic devices is an active research area for the past several decades. Multi-degree of freedom prosthetic hands controlled by surface electromyography signals (sEMG) of the forearm muscles have been developed to restore the lost limb functionality. A major objective of the current research is the development of intuitive prosthetic controllers for modern dexterous prosthetic hands. Current state of the art prosthetic controllers utilize pattern recognition (PR) of multi-channel sEMG signals for decoding the intended motion class. Despite decades of academic research, a clinically viable PR based prosthetic controller is yet to be realized. The PR based classifiers have resulted in more than 90% classification accuracy when evaluated under controlled laboratory conditions. However, the excellent laboratory performance of these classifiers is yet to result in their clinical acceptability and commercial availability. As a result, the currently available prosthetic devices are still based on non-intuitive binary or sequential digital control. There are several reasons for this academia-industry disparity. The unintended variations of sEMG signal characteristic due to several confounding factors, including arm position, adversely affects the performance of pre-trained PR based prosthetic controllers. Researchers have proposed fusion of auxiliary sensory information for classifier robustness against forearm positional variations. Sensor fusion techniques including accelerometer-mechanomyography (ACC-MMG), force-myography (FMG) and magnetic markers (MMs), have been reported with significant performance improvements. This study focused on the impact of arm position variations on the performance of PR-based forearm motion class decoders. A wearable data acquisition system was designed to acquire multi-channel sEMG and measure arm position using inertial measurement units. The performance of support vector machine (SVM) and linear discriminant analysis (LDA) classifiers was evaluated for training at static positions and for dynamic arm movements to characterize the adverse effects of arm position variation. Sensor fusion of sEMG and arm position data was evaluated to mitigate the arm position effect. A comparison of static multi-position training and dynamic arm movement training was carried out to suggest more pragmatic strategies for classifier training. The sensitivity of classifiers to motion class taxonomy was also evaluated. Sixteen motion classes categorized in distant-taxonomy and close-taxonomy groups were classified using LDA, Linear SVM (LSVM), Non-Linear SVM (NLSVM) and Multi-layer perception (MLP) classifiers. The results of the study have shown a significant dependence of LDA and LSVM classifiers and an insignificant dependence of the NLSVM and MLP classifiers on motion class taxonomy. The results suggest a more pragmatic selection of motion classes for realistic classifier performance evaluation.