Effect of the sEMG electrode (re)placement and feature set size on the hand movement recognition

Abstract: Repositioning of recording electrode array across repeated electromyography measurements may result in a displacement error in hand movement classification systems. In order to examine if the classifier re-training could reach satisfactory results when electrode array is translated along or rotated around subject's forearm for varying number of features, we recorded surface electromyography signals in 10 healthy volunteers for three types of grasp and 6 wrist movements. For feature extraction we applied principal component analysis and the feature set size varied from one to 8 principal components. We compared results of re-trained classifier with results from leave-one-out cross-validation classification procedure for three classifiers: LDA (Linear Discriminant Analysis), QDA (Quadratic Discriminant Analysis), and ANN (Artificial Neural Network). Our results showed that there was no significant difference in classification accuracy when the array electrode was repositioned indicating successful classification re-training and optimal feature set selection. The results also indicate expectedly that the number of principal components plays a key role for acceptable classification accuracy ~90 %. For the largest dataset (9 hand movements), LDA and QDA outperformed ANN, while for three grasping movements ANN showed promising results. Interestingly, we showed that interaction between electrode array position and the feature set size is not statistically significant. This study emphasizes the importance of testing the interaction of factors that influence classification accuracy and classifier selection altogether with their impact independently in order to establish guiding principles for design of hand movement recognition system. Data recorded for this study are stored on Zenodo repository (doi: 10.5281/zenodo.4039550).

• The paper investigates how sEMG electrode repositioning (translation/rotation) and feature set size (via PCA) affect hand movement classification accuracy using LDA, QDA, and ANN classifiers.
• Key findings indicate that classification accuracy is relatively robust to electrode repositioning but significantly influenced by feature set size, with more PCA components generally improving results.
• The study offers valuable insights for designing robust sEMG-based human-machine interfaces for applications like neuroprosthetics, suggesting future research on optimizing classifiers for variable electrode placements.

Analysis of sEMG Electrode Repositioning and Feature Set Size on Hand Movement Recognition

The research presented investigates the impact of surface electromyography (sEMG) electrode placement and feature size on the efficacy of hand movement classification systems. By utilizing a dataset from 10 healthy volunteers, the study explores how repositioning the electrode array, either through translation or rotation, influences classification accuracy across varying feature set sizes. The study employs Linear Discriminant Analysis (LDA), Quadratic Discriminant Analysis (QDA), and Artificial Neural Network (ANN) classifiers, thus providing a comprehensive understanding of classification accuracy under different conditions.

Key Findings and Methodology

1. Electrode Repositioning Effects: The study observes the influence of electrode array displacement on classification systems. The classifiers' ability to maintain accuracy is evaluated with electrodes translated and rotated on the forearm, and it is noted that classification accuracy remains relatively unaffected by these changes.
2. Feature Set Size Utilization: Principal Component Analysis (PCA) is harnessed to extract features from sEMG signals, varying from one to a fixed number of principal components. The results indicate that feature set size is a crucial determinant of classification accuracy, revealing that a larger number of components yields improved results.
3. Classifier Performance: The tested classifiers show different strengths based on the dataset. LDA and QDA outperform ANN in nine hand movements, whereas ANN is more promising in the grasping movements. The findings help outline the importance of selecting appropriate classifiers based on specific tasks.
4. Analysis and Results: It emerges that the number of principal components substantially influences classification success, typically achieving approximately 90% accuracy. The study particularly highlights that the interaction between electrode array position and feature set size remains statistically insignificant.
5. Data and Approach: The data are publicly accessible, providing transparency and allowing for further exploration. The methodology includes a rigorous protocol involving re-training of classifiers, ensuring robustness across diverse conditions.

Implications and Future Research

The findings analyze the relationship between electrode repositioning and feature set size. This is significant for sEMG-based interfaces, particularly in neuroprosthetic applications where sEMG reliability is critical. Potential developments may focus on optimizing classifiers and their calibration techniques to accommodate electrode repositioning by predicting or mitigating its effects on classification accuracy.

Furthermore, the study presents a possible framework for applying PCA-based feature selection in contexts with variable electrode placements, thus contributing valuable principles to the design of robust hand movement recognition systems. Future research could explore integrating various electrode positions into training and testing processes as well as apply findings to real-time systems and varied user groups beyond the initial 10-person sample.

In sum, this study lends insights into the optimization of classification accuracy under altered electrode distribution, offering essential contributions to the broader field of human-machine interface design and its practical implementations in assistive technologies.