An adaptive segmentation scheme based on recurring action potentials for sEMG controlled movement decoding.

An electromyography (EMG) controlled decoding system requires signal pre-processing, feature extraction, and classification as fundamental steps and requires high accuracy and minimum delay. The conventional system relies on the constant width segmentation scheme for feature extraction, which does not cover the complexities associated with the random behavior of EMG signals. An adaptive segmentation based on the repeating patterns of action potentials can be a promising solution. This work proposes a novel adaptive segmentation approach that captures the occurrence of these action potentials for segmentation and feature extraction. The proposed work is validated experimentally with 12 subjects performing eight different movements. Twenty-time domain features are extracted to verify the study. Linear Discriminant Analysis (LDA), k-nearest neighbor (kNN), and Decision Tree (DT) classifiers are used to observe the performance of the proposed scheme in terms of precision, recall, F1 score, and accuracy. The proposed method gives an average segmentation width of 124 ms across all subjects with 124 ± 5.4 (± 4.35 %) margin of error at 95 % confidence level. The average F1 score across all subjects for eight movements is 82.078 % for LDA, 81.51 % for kNN, and 80.81 % for DT classifiers. The 5-fold cross-validated accuracies for LDA, kNN, and DT classifiers are 78.3 %, 78.2 %, and 76.70 %, respectively. The calculated accuracies are compared with a constant width segmentation scheme with a window size of 200 ms. The t-test suggests significant improvement in the performance of the classifiers with the proposed method.