Enhancing motor imagery task recognition through local maximum synchro-squeezing transform and multi-domain features

• Published in: Biomedical signal processing and control Vol. 101; p. 107149
• Main Authors: Dovedi, Tanvi, Upadhyay, Rahul, Kumar, Vinay
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.03.2025
• Subjects: Brain-Computer Interface; Classification; Electroencephalogram; Feature Extraction; Local Maximum Synchro-Squeezing Transform; Non-negative Matrix Factorization; Feature Extraction; Non-negative Matrix Factorization; Local Maximum Synchro-Squeezing Transform; Brain-Computer Interface; Electroencephalogram; Classification
• ISSN: 1746-8094
• DOI: 10.1016/j.bspc.2024.107149

Motor imagery electroencephalogram signals play a crucial role in Brain-Computer Interface design that offers a means of communication for individuals with motor disabilities. The non-linear and non-stationary nature of electroencephalogram signals presents a significant challenge for accurate classification. The present work introduces a novel method for interpreting the motor imagery signals, aiming to enhance the classification performance. The proposed method is carried out in four methodological steps. In the first step, subject-specific electroencephalogram channel selection is performed to identify the most relevant channels and reducing the computational complexity of the method. The electroencephalogram signals obtained from the selected channels are decomposed into time–frequency coefficients using Local Maximum Synchro-Squeezing Transform in the second step. In the third step, the extracted time–frequency coefficients are treated using Non-negative Matrix Factorization for dimension reduction and effectively capturing patterns corresponding to the motor imagery tasks. Numerous multi-domain features are extracted and classified using seven different classifiers including Support Vector Machine, Linear Discriminant Analysis, K-Nearest Neighbour, Multi-Layer Perceptron, Extreme Gradient Boosting, Gradient Boosting, and Cat Boosting in the fourth step. The method is trained and tested on the BCI competition IV (2a) dataset. With an average classification accuracy reaching 98.44% for binary tasks and 90.00% for multi-class tasks, the method demonstrates robust performance, making it well-suited for real-time BCI systems.