Hierarchical strategy for sEMG classification of the hand/wrist gestures and forces of transradial amputees

Francesca Leone; Federico Mereu; Cosimo Gentile; Francesca Cordella; Emanuele Gruppioni; Loredana Zollo

doi:10.3389/fnbot.2023.1092006

Hierarchical strategy for sEMG classification of the hand/wrist gestures and forces of transradial amputees

Front Neurorobot. 2023 Mar 9:17:1092006. doi: 10.3389/fnbot.2023.1092006. eCollection 2023.

Authors

Francesca Leone¹, Federico Mereu^{1

2}, Cosimo Gentile^{1

2}, Francesca Cordella¹, Emanuele Gruppioni², Loredana Zollo¹

Affiliations

¹ Advanced Robotics and Human-Centred Technologies, Department at University Campus Bio-Medico of Rome, Rome, Italy.
² Istituto Nazionale Assicurazione Infortuni sul Lavoro (INAIL) Prosthetic Center, Vigorso, BO, Italy.

Abstract

Introduction: The myoelectric control strategy, based on surface electromyographic signals, has long been used for controlling a prosthetic system with multiple degrees of freedom. Several methods classify gestures and force levels but the simultaneous real-time control of hand/wrist gestures and force levels did not yet reach a satisfactory level of effectiveness.

Methods: In this work, the hierarchical classification approach, already validated on 31 healthy subjects, was adapted for the real-time control of a multi-DoFs prosthetic system on 15 trans-radial amputees. The effectiveness of the hierarchical classification approach was assessed by evaluating both offline and real-time performance using three algorithms: Logistic Regression (LR), Non-linear Logistic Regression (NLR), and Linear Discriminant Analysis (LDA).

Results: The results of this study showed the offline performance of amputees was promising and comparable to healthy subjects, with mean F1 scores of over 90% for the "Hand/wrist gestures classifier" and 95% for the force classifiers, implemented with the three algorithms with features extraction (FE). Another significant finding of this study was the feasibility of using the hierarchical classification strategy for real-time applications, due to its ability to provide a response time of 100 ms while maintaining an average online accuracy of above 90%.

Discussion: A possible solution for real-time control of both hand/wrist gestures and force levels is the combined use of the LR algorithm with FE for the "Hand/wrist gestures classifier", and the NLR with FE for the Spherical and Tip force classifiers.

Keywords: multi-DoFs control; pattern recognition; prosthetic control; real-time and offline performance; upper limb.

Grants and funding

This work was supported partly by the Italian Institute for Labor Accidents (INAIL) prosthetic center, through the WiFi-MyoHand (CUP: E59E19001460005) project, and partly by funding from the innovation program (Grant Agreement No. 899822, SOMA project).