Comprehensive Investigation on Features Selection using Statistical Tests for EMG Signal Classification

Gesture recognition is vital in applications such as bionic arms, robotics, virtual reality etc., development and deployments. EMG signals have gained traction for recognizing hand and finger gestures. This study focuses on investigating prominent features using statistical approaches. To achieve this a standard EMG dataset is made use off, comprising 36 participants with 8 unique hand gestures. The EMG signals were processed through three main stages: Firstly, feature extraction: temporal, frequency domain, logarithmic determinants, and logarithmic Teager-Kaiser Energy Operator (TKEO) features are extracted from the first-channel data, resulting in 40 feature dimension. Second, feature selection: statistical tests such as Analysis of Variance (ANOVA) and Kruskal-Wallis(KW) are employed to identify the most significant features and are ranked. Visualization of the significant and non-significant features reveals separability of each class. Classification is subsequently performed using Random Forest (RF) and Support Vector Machine (SVM) algorithms. Results demonstrate that the feature set combinations ranked by ANOVA, {LogTKEO} and {LogTKEO, Willison Amplitude} with RF classifier exhibits 0.85 and 0.86 accuracy respectively. However, features ranked by both ANOVA and KW have surpassed the accuracies in literature by 0.02 using single channel data.