XGBoost-Based Classification of Electromyography Signal for Muscle Contraction and Relaxation: A Correlation-Guided Feature Analysis Approach

Abstract

Analyzing electromyography (EMG) signals has become one of the important methods in biomedical engineering for researching neuromuscular function that is non-invasive, with applications such as robotic rehabilitation, prosthetics, and human-machine interface design. Despite significant improvements in EMG pattern recognition, robust classification is still constrained by feature redundancy, the models’ weak interpretability, and conventional statistical metrics that are ineffective at distinguishing between various physiological parameters. In an effort to overcome these constraints, this study proposes using a reliable Extreme Gradient Boosting (XGBoost) classifier to categorize segmented EMG signals as either contraction or relaxation states through a feature-correlation-driven method. Mean, standard deviation (SD), root mean square (RMS), maximum amplitude (MAX AMP), median frequency (MF), variance, skewness, and kurtosis are the eight time-domain and statistical metrics that were retrieved. Principal Component Analysis (PCA) assisted in visualizing class separation, while separately developed Pearson correlation heatmaps were used to investigate feature correlation. For the relaxation class and the contraction class, the XGBoost classifier obtained precision, recall, and F1-score values of 0.93 and 0.91, respectively. With an area under the receiver operating characteristic curve (AUC-ROC) score of 0.92, achieving an overall accuracy of 92% was evaluated on a 25-sample test set. Significant collinearity between amplified-based measurements was shown by correlation analysis, especially during muscular relaxation when the RMS-SD pair showed an almost perfect correlation of 1.00. Class segregation was validated by PCA along PC1 (61.5% variance explained) and PC2 (23.3%). Overall, wearable neuromuscular monitoring devices, EMG-driven rehabilitation interfaces, and smart prosthetic systems stand to benefit greatly from this approach’s reliable as well as computationally efficient real-time muscle activation monitoring methodology.