Bayesian Approach for Adaptive EMG Pattern Classification Via Semi-Supervised Sequential Learning

TL;DR

This paper introduces a Bayesian semi-supervised sequential learning method for adaptive EMG pattern classification, effectively maintaining accuracy over time despite signal variations due to electrode shifts and fatigue.

Contribution

It presents a novel Bayesian approach that updates classification models sequentially with pseudo-labels, improving robustness in EMG-based human-machine interfaces.

Findings

Reduces accuracy degradation over time

Outperforms conventional classification methods

Validated on experiments with six healthy adults

Abstract

Intuitive human-machine interfaces may be developed using pattern classification to estimate executed human motions from electromyogram (EMG) signals generated during muscle contraction. The continual use of EMG-based interfaces gradually alters signal characteristics owing to electrode shift and muscle fatigue, leading to a gradual decline in classification accuracy. This paper proposes a Bayesian approach for adaptive EMG pattern classification using semi-supervised sequential learning. The proposed method uses a Bayesian classification model based on Gaussian distributions to predict the motion class and estimate its confidence. Pseudo-labels are subsequently assigned to data with high-prediction confidence, and the posterior distributions of the model are sequentially updated within the framework of Bayesian updating, thereby achieving adaptive motion recognition to alterations in signal characteristics over time. Experimental results on six healthy adults demonstrated that the proposed method can suppress the degradation of classification accuracy over time and outperforms conventional methods. These findings demonstrate the validity of the proposed approach and its applicability to practical EMG-based control systems.