Semi-supervised style transfer mapping-based framework for sEMG-based pattern recognition with 1- or 2-DoF forearm motions. (July 2021)

• Record Type: Journal Article
• Authors: Kanoga, Suguru; Hoshino, Takayuki; Asoh, Hideki
• Is Part Of: Biomedical signal processing and control. Volume 68(2021)
• Publication Date: 2021-07
• Subjects: Surface electromyogram -- Transfer learning -- Style transfer mapping -- Covariate shift adaptation -- Pattern recognition; 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.2021.102817
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: A subject-to-subject transfer framework for surface EMG is proposed. 1- and 2-DoF datasets from 25 subjects using a wearable device are collected. The classifiers are trained with time-domain and autoregressive features. The proposed method exhibits superior performance over a conventional framework. The sEMG datasets and sample codes used are publicly available. Abstract: Background: The development of body-worn sensors has made it easier to measure surface electromyograms (sEMGs) from individuals and to employ sEMG-based interfaces for user-centric health monitoring, rehabilitation, human augmentation, and amusement. However, it remains difficult to measure large amounts of sEMG data from a user (target). Thus, the development of a subject-to-subject transfer framework that uses the information that is available from other people (source) is challenging work. Objectives: In this study, we propose a subject-to-subject transfer framework, which includes the following four steps: (i) the construction of individual support vector machines (SVMs) from the source subjects; (ii) the selection of effective classifiers for the target based on the individual SVM classification results; (iii) the linear projection of the target data into the source subject data using the semi-supervised style transfer mapping algorithm; and (iv) an ensemble strategy of class probabilities for the selected classifiers. Methods: To evaluate the performance of proposed framework, we collected 8-class, 1-DoF and 14-class, 2-DoF sEMG datasets that were acquired from the same 25 subjects using an eight-channel wearable device. The classifiers were trained with time-domain and autoregressive features. Results: Our proposed method exhibited superior performance on both datasets compared to a conventional transfer framework using covariate shift adaptation. The sEMG datasets and sample codes used in this study are publicly available at https://github.com/Suguru55/SS-STM_for_MyoDatasets .