Bagged tree ensemble modelling with feature selection for isometric EMG-based force estimation. (September 2022)

• Record Type: Journal Article
• Authors: Hajian, Gelareh; Behinaein, Behnam; Etemad, Ali; Morin, Evelyn
• Is Part Of: Biomedical signal processing and control. Volume 78(2022)
• Publication Date: 2022-09
• Subjects: High-density (HD) recording -- Surface electromyogram (EMG) -- Force estimation -- Isometric contraction -- Bagged Tree Ensemble (BTE); Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.104012
• Languages: English
• ISSNs: 1746-8094

Abstract

Abstract: EMG-based force estimation is crucial in applications, such as control of powered prosthetic and rehabilitation devices. Most previous studies focus on intra-subject force modelling. However, a generalized EMG-force estimation model, which is capable of estimating force across users, is needed for surgical and rehabilitation robotics. In this study, EMG signals are recorded from the long head and short head of the biceps brachii, and brachioradialis using 3 linear surface electrode arrays, during isometric elbow flexions, at different joint angles and forearm postures, while recording the induced force at the wrist. The recorded EMGs are pre-processed and segmented, and 336 time and frequency domain features are extracted from 21 EMG channels. We explore developing a model that can perform well across subjects, where Bagged Tree Ensemble (BTE) models are used to learn the non-linear, complex relationships between the EMG and force data. The BTE models are compared with several machine learning approaches. The BTE model performs best, giving an average normalized mean squared error (%NMSE) of 5.65±16.24%. To reduce the dimensionality of the feature space and improve force estimation performance, a novel feature selection technique, called modified sequential feature selection (MSFS) is implemented and compared to other commonly used feature selection methods. Results show that the MSFS algorithm outperformed other methods tested, significantly reducing the force estimation error. We also found that there was no effect of forearm posture and joint angle on force modelling accuracy, permitting the development of an isometric force estimation model generalized across participants and arm positions. Highlights: We utilize Bagged Tree Ensemble (BTE) models to accurately estimate force, using EMG features. Our study is one of the few works that attempts to estimate force in an inter-subject manner. A feature selection method, modified sequential feature selection (MSFS) is proposed. The MSFS reduces the feature set dimensionality and enhances the force modelling performance. No joint angle and forearm posture effect on force modelling performance is found.