نهمین کنگره نقشه برداری مغز

نمایش مقاله

Title:

Classification of Left- and Right-Hand Motor Imagery Using Electroencephalography (EEG) Signals

Subject:

Brain–Computer Interfaces (BCIs)

Presenter Author:
Name:	Yasaman Seyfi

Author:
Name 1:	Yasaman Seyfi - Corresponding - Presenter
Affiliation 1:	Department of Biomedical Engineering, Amirkabir University of Technology (Tehran Polytechnic), Tehran, Iran

Name 2:	Golnaz Baghdadi
Affiliation 2:	Department of Biomedical Engineering, Amirkabir University of Technology (Tehran Polytechnic), Tehran, Iran

Name 3:
Affiliation 3:

Name 4:
Affiliation 4:

Name 5:
Affiliation 5:

Name 6:
Affiliation 6:

Name 7:
Affiliation 7:

Name 8:
Affiliation 8:

Name 9:
Affiliation 9:

Name 10:
Affiliation 10:

Abstract

Original Research

introduction :

Electroencephalography-based motor imagery (EEG-MI) is a widely used paradigm in brain-computer interface systems. However, inter-subject variability may limit the generalization of classification models to unseen individuals. This study compared cross-subject and subject-specific approaches for classifying left- and right-hand motor imagery within a common analytical framework and examined patterns of model and feature selection.

Methods:

EEG data from the first 15 subjects of the publicly available EEG Motor Imagery dataset from GigaDB were analyzed. Recordings were obtained from 64 EEG channels at a sampling rate of 512 Hz. Preprocessing included filtering between 0.5 and 40 Hz and artifact removal using Extended Infomax independent component analysis. The EEG signals were subsequently segmented into epochs, and the 0.5–4.5 s interval following stimulus onset was used for feature extraction. Temporal- and frequency-domain features were extracted from each EEG channel, including amplitude-based and Hjorth parameters, absolute and relative power in the mu and beta bands, and normalized spectral entropy. To reduce dimensionality and limit the risk of overfitting, seven features were selected using the two-sided Mann–Whitney U test based exclusively on training data. Four classification algorithms—linear support vector machine (SVM), radial basis function SVM, random forest, and decision tree—were evaluated together with different hyperparameter configurations to identify the optimal model–hyperparameter combination based on validation balanced accuracy. In the cross-subject approach, model development and evaluation were performed using a nested leave-one-subject-out framework, in which the target subject was completely excluded from model development. In the subject-specific approach, 20% of each subject’s data was reserved as an independent final test set; feature selection, classifier selection, and hyperparameter optimization were performed on the remaining 80% using nested cross-validation with five outer folds and four inner folds. Balanced accuracy was used as the primary criterion for model selection and performance evaluation.

Results:

Mean subject-level balanced accuracy was 51.9 ± 5.9% for the cross-subject approach and 83.3 ± 14.7% for the subject-specific approach. The subject-specific approach achieved higher balanced accuracy than the cross-subject approach for all 15 subjects. The optimal classifier and final feature subset varied among subjects, suggesting substantial inter-subject heterogeneity in model and feature selection. Despite this variability, several absolute beta-band power features, particularly those derived from AF7, AF4, and AF8, were repeatedly selected in multiple subjects, suggesting the presence of partially shared discriminative EEG patterns.

Conclusion:

Inter-subject variability remains an important challenge for the generalization of EEG-based motor imagery classification models. In this study, the subject-specific approach achieved higher balanced accuracy than the cross-subject approach. The observed differences in classifier selection and feature selection highlight the importance of subject-specific optimization in EEG-based brain-computer interface systems. These findings suggest that incorporating individual characteristics may improve the performance of motor imagery classification models.

Keywords:

Brain-Computer Interfaces; Electroencephalography; Motor Activity; Machine Learning