Reduced Daily Recalibration of Myoelectric Prosthesis Classifiers Based on Domain Adaptation

• Published in: IEEE journal of biomedical and health informatics Vol. 20; no. 1; pp. 166 - 176
• Main Authors: Liu, Jianwei, Sheng, Xinjun, Zhang, Dingguo, He, Jiayuan, Zhu, Xiangyang
• Format: Journal Article
• Language: English
• Published: United States IEEE 01.01.2016; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Adaptation; Adaptation models; Amputees; Artificial Limbs; Calibration; Classification; Classifiers; Data models; Discriminant Analysis; domain adaptation; Electromyography - methods; Electromyography - standards; Feature extraction; Female; Humans; linear discriminant analysis; Male; Middle Aged; Pattern recognition; Pattern Recognition, Automated - methods; patternrecognition; polynomial classifier; Polynomials; Prostheses; Prosthetics; Reuse; Signal Processing, Computer-Assisted - instrumentation; Surface electromyography (sEMG); Surgical implants; Training; Usability; Vectors; linear discriminant analysis (LDA); polynomial classifier (PC); surface electromyography (sEMG); Domain adaptation (DA); pattern recognition
• ISSN: 2168-2194; 2168-2208; 2168-2208
• DOI: 10.1109/JBHI.2014.2380454

Control scheme design based on surface electromyography (sEMG) pattern recognition has been the focus of much research on a myoelectric prosthesis (MP) technology. Due to inherent nonstationarity in sEMG signals, prosthesis systems may need to be recalibrated day after day in daily use applications; thereby, hindering MP usability. In order to reduce the recalibration time in the subsequent days following the initial training, we propose a domain adaptation (DA) framework, which automatically reuses the models trained in earlier days as input for two baseline classifiers: a polynomial classifier (PC) and a linear discriminant analysis (LDA). Two novel algorithms of DA are introduced, one for PC and the other one for LDA. Five intact-limbed subjects and two transradial-amputee subjects participated in an experiment lasting ten days, to simulate the application of a MP over multiple days. The experiment results of four methods were compared: PC-DA (PC with DA), PC-BL (baseline PC), LDA-DA (LDA with DA), and LDA-BL (baseline LDA). In a new day, the DA methods reuse nine pretrained models, which were calibrated by 40 s training data per class in nine previous days. We show that the proposed DA methods significantly outperform nonadaptive baseline methods. The improvement in classification accuracy ranges from 5.49% to 28.48%, when the recording time per class is 2 s. For example, the average classification rates of PC-BL and PC-DA are 83.70% and 92.99%, respectively, for intact-limbed subjects with a nine-motions classification task. These results indicate that DA has the potential to improve the usability of MPs based on pattern recognition, by reducing the calibration time.