社団法人 電子情報通信学会 THE INSTITUTE OF ELECTRONICS, INFORMATION AND COMMUNICATION ENGINEERS 信学技報 TECHNICAL REPORT OF IEICE. Continuous Estimation of Finger Joint Angles Using Inputs from an EMG-to-Muscle Activation Model Jimson G. NGEO † , Tomoya TAMEI † , and Tomohiro SHIBATA † † Graduate School of Information Science, Nara Institute of Science and Technology 8916–5 Takayama, Ikoma, Nara, 630–0192 Japan Abstract Surface electromyography (sEMG) signals are often used in many robot and rehabilitation applications because these reflect the motor intention of users. However, inherent problems such as electromechanical delay are present in such applications. Here, we present a method to estimate finger joint angles using a neural network with inputs obtained from an EMG-to-Activation model which parameterizes this delay. Our results show overall root-mean-square errors of 5-12% be- tween the predicted and actual joint angles. We also show results when the proposed muscle activation input is used compared to using features used by other related studies. Finally, we compare the use of a neural network to a Gaussian Process, which is a popular nonparametric Bayesian regressor that could efficiently give better prediction in this setting. Key words Surface Electromyography (sEMG), Muscle Activation Model, Finger Joint Angles, Neural Networks 1 Introduction In the coming years, human assistive and tele-manipulation tech- nologies are expected to play a significant role in improving the lives of the aged, handicapped and injured. For instance, assistive robotic devices and brain-machine interfaces are developing tech- nologies that have the potential to give tireless support, enabling persons with impairments and injuries to achieve more function and mobility. Given the dominant role of the upper limb, where many everyday functional tasks are achieved by our hands, providing a means to accurately replicate these motions can greatly improve current support devices, particularly, in the development of devices and applications, such as hand and finger exoskeletons, that can aid in hand rehabilitation. Tele-operated devices controlled by neural signals can give un- constrained and precise movement control in different environments [1]. Surface electromyogram (sEMG) signals are often used in prosthesis controls and rehabilitation support applications because these reflect the motor intention of a user prior to the actual move- ments [2]. These signals not only provide little delay when used in human interfaces, but have also been shown to represent muscle tension and joint positions very well. In the field of replicating motions, discrete classification of hand gestures have been successful, reaching a decoding accuracy of above 95% and classifying to up to more than 20 gestures [3]. How- ever, natural hand movements are not limited to discrete gestures but are continuous and coordinated. As an initial step, our research aims to predict multiple finger joint-angles simultaneously and continu- ously from inputs based on an EMG-to-Muscle activation model. Related studies have shown that it is possible to extract fine fin- ger movement information contained in sEMG signals. Afshar and Matsuoka [1] were able to estimate the index finger joint angles from fine-wire EMG embedded inside seven muscles that control the index finger. Similarly, Shrirao et al.[4] were able to decode one index finger joint angle but from surface EMG signals. The types of finger motion involved in their study were periodic flexion- extension motion at three different frequencies, and they evaluated different types of neural network to predict the joint angle. In a more recent development, Smith et al. [5] were able to asynchronously de- code individual metacarpophalangeal (MCP) joint angles of all five fingers while moving one finger at a time. Their study extracted sEMG time-domain features from general muscle locations avail- able to transradial amputees and used these features as input to a feed-forward neural network to estimate the joint angles. However in the previous studies [4] [5], a time delay between the onset of the sEMG signal and exerted movement was present and observed. This time delay is called hysteresis or electromechani- cal delay (EMD). It is often compensated by either manually re- aligning the sEMG to the joint angle data before any processing is done or by introducing time-delay lines, which makes use of all the immediate and past values of the sEMG, greatly increasing the number of inputs and as well as parameters of the regressor used. EMD can vary depending on many different factors such as mus- cle shortening velocity, type of muscle fiber, and fatigue[6]. In our method, we introduce this delay as a parameter, by using an EMG- to-Muscle Activation Model, which is determined along with other parameters through optimization. Here, we investigate the use of muscle activation as input in estimating both periodic and nonpe- riodic flexion and extension movement of all five finger joint an- gles. We attempt to predict the angular position of each finger joint, namely, the metacarpophalangeal (MCP), proximal interphalangael (PIP) and the distal interphalangeal (DIP) joints. —1—