EEG-Driven Arm Movement Decoding: Combining Connectivity and Amplitude Features for Enhanced Brain–Computer Interface Performance

• Published in: Bioengineering (Basel) Vol. 12; no. 6; p. 614
• Main Authors: Darvishi, Hamidreza, Mohammadi, Ahmadreza, Maghami, Mohammad Hossein, Sadeghi, Meysam, Sawan, Mohamad
• Format: Journal Article
• Language: English
• Published: Switzerland MDPI AG 04.06.2025; MDPI
• Subjects: Accuracy; Algorithms; Amplitudes; Arm; Artificial neural networks; Bandpass filters; Biochips; Brain; Brain research; brain–computer interface; Computer applications; Control systems; Decoding; EEG; Electroencephalography; Electromyography; filter bank common spatial patterns; Filter banks; Fourier transforms; Frequencies; Frequency dependence; Human-computer interface; Implants; Kinematics; Muscle function; neural network; Neural networks; Neural prostheses; Neurophysiology; phase-locking value; Prostheses; Prosthetics; Resampling; Robust control; Root-mean-square errors; electroencephalography; filter bank common spatial patterns; electromyography; brain–computer interface; movement decoding; phase-locking value; neural network; ReliefF; feature selection
• ISSN: 2306-5354; 2306-5354
• DOI: 10.3390/bioengineering12060614

Brain–computer interfaces (BCIs) translate electroencephalography (EEG) signals into control commands, offering potential solutions for motor-impaired individuals. While traditional BCI studies often focus solely on amplitude variations or inter-channel connectivity, movement-related brain activity is inherently dynamic, involving interactions across regions and frequency bands. We propose that combining amplitude-based (filter bank common spatial patterns, FBCSP) and phase-based connectivity features (phase-locking value, PLV) improves decoding accuracy. EEG signals from ten healthy subjects were recorded during arm movements, with electromyography (EMG) as ground truth. After preprocessing (resampling, normalization, bandpass filtering), FBCSP and multi-lag PLV features were fused, and the ReliefF algorithm selected the most informative subset. A feedforward neural network achieved average metrics of: Pearson correlation 0.829 ± 0.077, R-squared value 0.675 ± 0.126, and root mean square error (RMSE) 0.579 ± 0.098 in predicting EMG amplitudes indicative of arm movement angles. Analysis highlighted contributions from both FBCSP and PLV, particularly in the 4–8 Hz and 24–28 Hz bands. This fusion approach, augmented by data-driven feature selection, significantly enhances movement decoding accuracy, advancing robust neuroprosthetic control systems.