Variable Transformation for sEMG Pre-Processing: Applications and Analysis in Predicting Human Upper Limb Motion Intentions

Abstract

Surface electromyography (sEMG) signals hold considerable promise for predicting upper limb motion, particularly in applications related to human-machine interaction and rehabilitation. While extensive research has been conducted on sEMG signal analysis and pre-processing techniques, relatively little attention has been paid to addressing inter-subject variability in motion prediction using deep learning models. This study presents a comparative evaluation of six distribution transformation techniques, applied during pre-processing to account for subject-specific differences in sEMG signal distributions for upper limb intention prediction. The experimental results indicate that normalization choice significantly affects model performance, with techniques such as Yeo-Johnson leading to the highest accuracy improvement of up to 12.05% in Bi-LSTM prediction models. These findings underscore the importance of selecting appropriate personalized distribution transformation (PDT) strategies to enhance the generalizability and effectiveness of sEMG-based motion intention prediction.