Real-Time Classification of Hand Movements from Surface Electromyography Signals Using Feature Engineering and an Ensemble of Machine-Learning Classifiers for Rehabilitation

Turimov Dilmurod; Kilichev Dusmurod; Saydirasulov Norkobil; Mukhamadiyev Abdinabi

doi:10.65164/46rx7x24

Authors

Turimov Dilmurod Department of Computer Engineering, Gachon University, Sujeong-gu, Seongnam-si 461-701, Gyeonggi-do, Republic of Korea, Department of Control Theory and Information Security, Samarkand State University named after Sharof Rashidov, Samarkand 140104, Uzbekistan Author
Kilichev Dusmurod Department of Information Security, National University of Uzbekistan, Tashkent 100174, Uzbekistan Author
Saydirasulov Norkobil Department of Computer Engineering, Gachon University, Sujeong-gu, Seongnam-si 461-701, Gyeonggi-do, Republic of Korea, Department of Technologies of Mobile Communication, Tashkent University of Information Technologies Named After Muhammad Al-Khwarizmi, Tashkent 100200, Uzbekistan Author
Mukhamadiyev Abdinabi Department of Computer Engineering, Gachon University, Sujeong-gu, Seongnam-si 461-701, Gyeonggi-do, Republic of Korea Author

DOI:

https://doi.org/10.65164/46rx7x24

Keywords:

surface electromyography; hand-movement classification; feature engineering; machine learning; ensemble classifier; real-time; myoelectric prosthesis; rehabilitation; spinal cord injury.

Abstract

Myoelectric control of prostheses and biosignal-driven rehabilitation require handmovement recognition that is simultaneously accurate, computationally light, and fast enough for
real-time operation on embedded hardware. This paper presents an end-to-end methodology for
classifying object-interaction hand movements from surface electromyography (sEMG) signals.
Four contributions are combined into a single deployable pipeline: (i) a real-time active-potential
onset/offset detection algorithm based on rectified-signal variance thresholding (threshold = μ +
hσ); (ii) an optimal, low-dimensional feature vector obtained through feature engineering in the
time, frequency and time-frequency domains; (iii) a real-time algorithm for automatically building
the training set during signal acquisition; and (iv) an ensemble classifier that fuses k-Nearest
Neighbours, Support Vector Machine and Random Forest decisions with a confidence gate and
temporal confirmation. Two datasets were recorded - a single-channel BITalino dataset (20
subjects, 4 movements) and a five-channel BTS FreeEMG dataset (42 subjects, 7 movements). The
proposed onset detector achieved the lowest mean error (15.45 samples) among six competing
methods.