부산대학교 도서관

Electroencephalogram (EEG) signal processing plays a pivotal role in deciphering profound neurophysiological insights from seemingly inconspicuous neural signals, enabling the realization of practical BCI applications. This paper delves into an innovative and sophisticated EEG signal processing model, which employs three distinct combinations of EEG electrodes, four advanced feature extraction methods, and four cutting-edge classification algorithms, in conjunction with an adaptive multivariate chirp mode decomposition (AMCMD) for the analysis of motor imagery EEG signals. The feasibility of this novel approach is corroborated through its successful application to a sizable GigaDB dataset comprising 52 participants, as well as the challenging BCI competition III datasets IVa and IVb. The results unequivocally demonstrate that the seamless integration of the AMCMD mechanism with an 18-electrode combination, spectral features, and a multilayer neural network classifier engenders remarkably robust and discriminating classification outcomes. The classification accuracy for dataset IVa and IVb subjects attains an impressive pinnacle, reaching 99.70%, 99.88%, 99.88%, 99.87%, 100%, and 94.98%, respectively. Furthermore, when scrutinizing the GigaDB dataset, the average classification accuracy, sensitivity, specificity, and f1score exhibit commendable values of 82.71%, 82.76%, 83.04%, and 82.80%, respectively. These salient findings, juxtaposed against the backdrop of earlier studies, decisively manifest a striking 15.4% enhancement in average classification accuracy, irrefutably affirming the exceptional prowess of the proposed model. Consequently, the compelling evidence culminates in the conclusion that the adaptive AMCMD method stands resolutely robust, adept, and universally adaptive in its ability to proficiently classify EEG signals, deftly circumventing subject-to-subject variability across diverse and extensive datasets.