Multimodal Fusion Approach Based on EEG and EMG Signals for Lower Limb Movement Recognition

• Published in: IEEE sensors journal Vol. 21; no. 24; pp. 27640 - 27650
• Main Authors: Al-Quraishi, Maged S., Elamvazuthi, Irraivan, Tang, Tong Boon, Al-Qurishi, Muhammad, Parasuraman, S., Borboni, Alberto
• Format: Journal Article
• Language: English
• Published: New York IEEE 15.12.2021; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Classifiers; Correlation analysis; EEG; Electrodes; Electroencephalography; Electromyography; EMG; Fatigue; fusion; Joints; Motion perception; Multimodal; Muscles; Muscular fatigue; Pattern recognition; pattern recognition ankle joint movements; Recognition; Sensors; Statistical analysis
• ISSN: 1530-437X; 1558-1748
• DOI: 10.1109/JSEN.2021.3119074

In this study, the fusion of cortical and muscular activities based on discriminant correlation analysis DCA) is developed to recognize bilateral lower limb movements. Electromyography (EMG) and electroencephalography (EEG) signals were concurrently recorded from 28 healthy subjects while performing various ankle joint movements. The two types of biosignals were fused at feature level, and five different classifiers were used for the purpose of movement recognition. The performance of the classifiers with multimodal and single modality data were assessed with five different sampling window sizes. The results demonstrated that the use of a multimodal approach results in an improvement of the classification accuracy with a linear discriminator analysis classifier (LDA). The highest recognition accuracy was 96.64 ± 4.48% with a window size of 250 sample points, in contrast with 89.99 ± 7.94% for EEG data alone. Furthermore, the multimodal fusion based on DCA was validated with fatigued EMG signal to investigate the robustness of the fusion technique against the muscular fatigue. In addition, the statistical analysis result demonstrates that the proposed fusion approach provides a substantial improvement in motion recognition accuracy 96.64 ± 4.48% (p < 0.0001) compared to method based on a single modality.