Lorenzo Masia

Machine Learning for Human Motion Intention Detection using Myoelectric Interface



Myoelectric based interfaces offer a wide variety of applications in the area of prostheses, orthosis and tele-manipulation. EMG signals can be considered as a useful proxy to the signals originating from the brain. Myoelectric controlled interfaces (MCI) are also largely used since they are non-invasive, and are quite practical in the control of powered prosthetics and exoskeletons. EMG signals provide a much better understanding of the intended motion, since they are only generated when the muscle is stimulated, which directly implies the intended motion. Also EMG signals can be used to describe co-contraction patterns, i.e. adjustment of musculoskeletal impedance.

Feature extraction techniques help in characterizing underlying structure in the EMG signal. The set of features used, contribute to a great extent in the performance of any learning algorithm. There are broadly two types of approaches: EMG features, which looks into individual muscle activity like time, frequency and time-frequency domain features, and Synergy features, which looks into information from multiple muscle activities. Synergy features extract coordination patterns across the multiple EMG channels, by means of time-variant or time-invariant synergies.

Apart from feature extraction techniques, the type and amount of training data used also does affect the performance of the decoder. Incorporating dynamically varying data and including multiple limb positions have shown to improve decoding performance, but causes an increase in the training data-set and ultimately leading to higher computational burden. Extreme learning machine (ELM) is a relatively new supervised learning algorithm and represents a single-hidden layer feed-forward neural network (SLFNN). The learning rates are significantly higher than the traditional back-propagation based learning machines, and it provides an efficient solution to generalized feed-forward networks. ELM offers faster rates of training, less degree of intervention and ease of implementation.


train test2

Figure: A) Represents the offline training phase, where the graphical user interface instructs the subject to perform specific movements for specific period of time. The EMG data is then used to build the decoding model which is specific for each subject. B) The online testing phase, where EMG signals are decoded in real-time, and facilitates the movement of the virtual avatar.

The main focus of this paper is to compare and contrast the performance of ELM using the two approaches EMG features (more specifically time-domain features) and Synergy features; this study will focus on the differences between the two approach in decoding shoulder and elbow motions.

motion pose3

Figure: The five different motion-classes and the rest pose which represents the sixth class. The movement starts from the rest pose and transitions to the target pose as indicated in the figure, and the motion is designed to follow a minimum-jerk trajectory.


One main drawback of this decoding strategy is that it involves sequential classification. It is not possible to use more than one type of movements. The next important focus of this research is implementing simultaneous decoding strategy. This would prove to be very important in the scenario involving hand and wrist motions, thereby increasing the dexterity and usability of the prostheses and orthoses.


Antuvan, C.W., Bisio, F., Marini, F., Yen, S.C., Cambria, E. and Masia, L., 2016. Role of Muscle Synergies in Real-Time Classification of Upper Limb Motions using Extreme Learning Machines. Journal of NeuroEngineering and Rehabilitation13, pp.1-15.

Antuvan, C.W., Bisio, F., Cambria, E. and Masia, L., 2015, August. Muscle synergies for reliable classification of arm motions using myoelectric interface. In 2015 37th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC) (pp. 1136-1139). IEEE.

Antuvan, C.W., Bisio, F., Cambria, E. and Masia, L., 2015, August. Discrete classification of upper limb motions using myoelectric interface. In 2015 IEEE International Conference on Rehabilitation Robotics (ICORR) (pp. 451-456). IEEE.



Pattern recognition-based control scheme is generally adopted to train decoders in predicting movements pertaining to unseen data. The drawback of pattern recognition algorithms (especially in the context of discrete classification), is that the decoding is sequential: meaning that the decoder can only predict one output at a time, even if the subject is performing more than one movement simultaneously. Simultaneous movements are a natural and efficient way by which humans perform activities of daily living, and as such it is important to incorporate this aspect into myoelectric-based motion decoding.

                                       exp setup

Figure: Online Testing Phase: EMG sensors capture the electrical activity, and the raw EMG signals areprocessed in real-time and outputs the predicted motion-class, which then controls the motion of the avatar.

We propose a unique classification strategy designed to incorporate simultaneous decoding of motions involving wrist (wrist flexion/extension) and hand (hand open/close). The performance of the decoding algorithm is evaluated for online control of a virtual hand model. While simultaneous motion decoding has been attempted before, our contribution is in developing a subject-independent classifier, which is based on the relative EMG activity of the various muscles utilized in the experiment.                           

                                     handpose muscles

Figure: The various movements performed by the virtual avatar is indicated on the left, and the various muscles used and their electrode psoitions are shown on the right.



 Antuvan, C.W., Yen, S.C. and Masia, L., 2016, June. Simultaneous classification of hand and wrist motions using myoelectric interface: Beyond subject specificity. In 2016 6th IEEE International Conference on Biomedical Robotics and Biomechatronics (BioRob) (pp. 1129-1134). IEEE.



IEEE BIOROB 2016 (clickhere)



  • L. Masia, Nanyang Tech. University
  • S. Kukreja, National University of Singapore



New Journal on Rehabilitation Technology

Journal of Rehabilitation and Assistive Technologies Engineering (click here)

Lorenzo MASIA (Associate Editor)



Journal Topics

Research Topic (click here)





International Conference on Rehabilitation Robotics

Singapore 11th-14th August 2015 Hosted by NTU

General Chair: Wei Tech ANG

Program Chair: Lorenzo MASIA



Lorenzo MASIA Editor of Biorob 2014



Hot the NEW WrisBot




visit the robot report website for a comprehensive view of robotic business worldwide and other great info!