Learning-based EEG rhythm analysis for the enhancement of motor imagery-based brain-computer interface performance

One of the major challenges in achieving stability and adaptability in motor imagery brain-computer interfaces (MI-BCI) is the individual variability in motion-related neural activity. This paper presents a novel approach to motor imagery electroencephalogram (MI-EEG) data analysis and MI-BCI optimization, making the following contributions: (1) We introduce EhythmNet, a compact one-dimensional convolutional neural network (CNN) designed to capture the inherent frequency features in MI-EEG data. (2) We propose using kernel coefficients derived from EhythmNet to construct a temporal bandpass filter within the conventional MI-BCI framework, where independent component analysis (ICA) is employed as the spatial filter. Experimental results consistently show that replacing the conventional bandpass filter (CBF) with a kernel-based bandpass filter (KBF) significantly improves stability and MI classification performance (KBF-ICA: 82.3 +/- 2.7 %, CBF-ICA: 72.4 +/- 7.5 %, EEGnet: 74.5 +/- 6.3). (3) To optimize MI-BCI, we advocate for integrating KBF with the ICA spatial filter. Our experimental findings clearly demonstrate that the KBF-ICA approach outperforms traditional methods, such as common spatial pattern (CSP) and EEGnet, particularly in cross-subject testing scenarios (KBF-ICA: 70.8 %, KBFCSP: 50.6 %, EEGnet: 45.1 %). Furthermore, with optimized EEG channel selection, the recognition rate of KBFICA can reach 91.8 %, whereas the other two methods, CSP and EEGnet, do not exhibit significant performance improvement through channel optimization. In conclusion, we highlight the potential of combining traditional MI-BCI decoding models with deep learning techniques to address current performance limitations in MI-BCIs. This integration holds significant theoretical and practical implications, providing a promising direction for future advancements in MI-BCI technology.