Prototype Reduction on sEMG Data for Instance-based Gesture Learning towards Real-time Prosthetic Control

Abstract

Current systems of electromyographic prostheses are controlled by machine learning techniques for gesture detection. Instance-based learning showed promising results concerning classification accuracy and robustness without explicit model training. However, it suffers from high computational demands in the prediction phase, which can be problematic in real-time scenarios. This paper aims at combining such learning schemes with the concept of prototype reduction to decrease the amount of data processed in each prediction step. First, a suitability assessment of state-of-research reduction algorithms is conducted. This is followed by a practical feasibility analysis of the approach. For this purpose, several datasets of signal classes from exerting specific gestures are captured with an eight-channel EMG armband. Based on the recorded data, prototype reduction algorithms are comparatively applied. The dataset reduction is characterized by the time needed for reduction as well as the possible data reduction rate. The classification accuracy when using the reduced set in crossvalidation is analyzed with an exemplary kNN classifier. While showing promising values in reduction time as well as excellent classification accuracy, a reduction rate of over 99% can be achieved in all tested gesture configurations. The reduction algorithms LVQ3 and DSM turn out to be particularly convenient. 1 MOTIVATION AND RELATED WORK The k nearest neighbour classification technique (kNN) as an exemplary instance-based machine learning scheme has shown several advantageous properties in the context of gesture recognition (intent detection) based on surface electromyography (sEMG) signals for prosthetics. In preliminary studies, kNN showed promising results in classification accuracy, generalizability, as well as user study success rate (Cipriani et al., 2011; Geethanjali, 2015; Tello et al., 2013; Khushaba et al., 2016; Sziburis et al., 2020). Moreover, it turned out to perform well in terms of robustness, i. a., against sampling frequency variation (Chen et al., 2017) and electrode shift (Li et al., 2016). It is characterized by a comparably low implementation complexity. Furthermore, instance-based learna https://orcid.org/0000-0002-7741-1276 b https://orcid.org/0000-0002-0840-5155 c https://orcid.org/0000-0001-5110-6823 ing schemes provide the benefits of no explicit mathematical model generation, and incrementality, i. e. the possibility to extend the dataset by new samples at any time with them being equally considered. While these characteristics speak in favour of embedded applicability in the context of wearable realtime systems, an important drawback of instancebased learning schemes is the necessity of comparing new arriving data instances (samples) in the prediction phase to all already stored ones. The needed iterations over all samples lead to potentially computationally intense operations, depending on the amount of data, i. e. the samples to be iterated in every prediction step (Sziburis et al., 2020). For this reason, this work analyzes the suitability of concepts to reduce the computational effort during prediction in instance-based learning schemes. Although no explicit model training takes place, the data stored in memory is referred to as training data in this paper. Two main approaches to improve the performance in this regard can be pointed out. The first possibility Sziburis, T., Nowak, M. and Brunelli, D. Prototype Reduction on sEMG Data for Instance-based Gesture Learning towards Real-time Prosthetic Control. DOI: 10.5220/0010327502990305 In Proceedings of the 14th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2021) Volume 4: BIOSIGNALS, pages 299-305 ISBN: 978-989-758-490-9 Copyright c © 2021 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved 299 is the utilization of memory-efficient, optimized data structures, e. g. “ball-tree data structures, hashing” (Kusner et al., 2014). The second approach is data reduction and aims at decreasing the amount of signal data gathered during training and stored in memory. It can be principally applied in a horizontal (feature space) and in a vertical dimension (sample space). Aside from that, there are techniques using approximations, such as Large Margin Nearest Neighbour (Kusner et al., 2014) for kNN. Horizontal data reduction is limited by the high variance of EMG signals. Nevertheless, the concept of feature selection (or horizontal thinning) has been applied in the context of pattern-recognitionbased prosthetic control for datasets of high feature space dimensions, e. g. in the form of biologically inspired methods such as genetic algorithms and particle swarm optimization (Purushothaman, 2016). Further concepts of horizontal data reduction are feature discretization, projection and positioning (Kusner et al., 2014). These come along with dimensionality reduction algorithms, e. g. PCA (Güler and Koçer, 2005) and adaptions (Nagata et al., 2005) as well as variants of LDA (Negi et al., 2016). The concept of vertical data reduction is mainly referred to as prototype reduction. This paper will examine techniques of this group in the context of EMG gesture data. Prototype stands for data instance or sample. However, it also indicates that it refers to specific instances that represent a larger amount of instances to a certain extent. Garcı́a et al. (Garcı́a et al., 2012) and Triguero et al. (Triguero et al., 2012) provide an overview of a variety of prototype reduction algorithms proposed. The methods are divided into two groups: On the one hand prototype selection (selecting a subset of instances from the existing ones stored in memory, also called vertical thinning), and on the other hand prototype generation (creating new instances based on the existing ones to represent the whole dataset). 2 CONCEPTUAL APPROACH The presented approach consists of reducing the computational effort of prediction steps by decreasing the number of samples within the gesture dataset. For this purpose, the term reduction rate describes the number of samples finally stored in memory relative to the number of originally recorded samples. 2.1 Theoretical Requirements First, an assessment of the variety of algorithms reviewed in (Garcı́a et al., 2012) for prototype selection and (Triguero et al., 2012) for prototype generation will be conducted, regarding the suitability of each method. Suitable methods should not influence the incrementality of the applied instance-based classifier and provide a possibility to specify the number of prototypes in the final set or accordingly the reduction rate beforehand. The latter characteristic will be called size determinism in the following. Moreover, there is the requirement of real-time capability which refers to the prediction phase of gesture recognition, since this phase takes place online and continuously decides on user satisfaction. The real-time property is guided by timing determinism, i. e. executing an identical number of operations per prediction. This can be achieved by establishing the same number of stored instances to be considered in each prediction step. To guarantee that, size determinism is inherently necessary. In order to provide fast reaction times, the number of stored samples should be as low as possible. Additionally, by providing size determinism, deterministic memory demands are facilitated. In the training phase, in contrast, sample capturing and offline calculations take place, which are not meant to be applied in real-time. Therefore, it is not particularly needed to cope with real-time requirements in this phase. However, training computations should progress as fast as possible to avoid delays for the user between training and utilizing the gesture prediction system. 2.2 Practical Methods After their theoretical assessment, promising algorithms will be practically evaluated in experiments on datasets of the linear envelope of rectified EMG signals. They consist of samples captured when exerting multiple sets of varying gesture configurations with an eight-channel state-of-the-art armband positioned on the forearm. With that, one sample is composed of eight 32-bit floating-point values. Within one dataset, four repetition blocks are alternatingly recorded for each of the gestures, whereat each repetition block consists of 400 samples (i. e. two seconds offline data sample recording at a capturing rate of 200 Hz) per gesture. Overall, three such datasets with slightly differing sensor positionings are recorded per gesture configuration. The following differing gesture configurations are selected in order to test the reduction algorithms’ poBIOSIGNALS 2021 14th International Conference on Bio-inspired Systems and Signal Processing