همایش ملی بیوانفورماتیک ایران

صفحه اصلی / 4th international edition and 13th Iranian Conference on Bioinformatics

A Clustering-Based Method for Preserving Manifold Structure in EEG Signals Classification

نویسندگان :

Shermin Shahbazi¹ Majid Ramezani²

1- دانشگاه زنجان 2- دانشگاه تحصیلات تکمیلی علوم پایه زنجان

کلمات کلیدی :

BCI،EEG_signal_classification،Riemannian_manifold،clustering_based_classification،Ensemble_modeling

چکیده :

Accurate classification of EEG signals is essential for brain-computer interface (BCI) and neuroprosthetic applications, enabling precise interpretation and control. Traditional EEG classification methods often ignore the non-Euclidean manifold structure in EEG data, which distorts relationships between signals and reduces classification accuracy (Lotte et al., 2018). Preserving this manifold information is crucial for capturing the true geometry of EEG signals, which Euclidean-based approaches inadequately address (Fu et al., 2024; Tibermacine et al., 2024). To overcome this limitation, this study introduces Clustering-Based Methods for Preserving Manifold Structure in EEG Signal Classification, a method that maintains Riemannian manifold geometry by preserving curvature-related information, ensuring essential non-Euclidean relationships are retained for accurate classification. The methodology comprises four phases: in feature engineering phase, covariance matrices and radial basis function (RBF) kernels are used to capture linear and non-linear relationships among EEG channels, enriching the data with a comprehensive view of brain activity (Davis and Curriero, 2019; Uehara et al., 2017). In data clustering phase, the feature-enhanced data is projected into a Riemannian manifold space, where clustering is performed using a novel k-means algorithm. This algorithm employs a metric combining Riemannian distance and tangent plane slope, preserving intrinsic structural relationships through locally sensitive clustering. In the dimensionality reduction phase, clustered data points are projected onto the tangent space of the manifold, reducing dimensionality for simplified representation in subsequent classification. This projection retains essential local structures, reduces noise, and prevents overfitting while preserving manifold-related information (Gao et al., 2021). The tangent space provides a linear approximation of the manifold, facilitating efficient handling of high-dimensional data in a lower-dimensional Euclidean context (Djebra et al., 2022). Finally, in the classification phase, a support vector machine (SVM) classifier is applied to the dimensionally reduced, clustered data. By leveraging preserved geometric features, the SVM enables efficient, accurate classification, suitable for real-time BCI and neuroprosthetic applications. The method’s effectiveness is validated on the BCI Competition IV dataset 2a (Tangermann et al., 2012), which includes 25 channels (22 EEG and 3 EOG) sampled at 250 Hz, recorded from nine subjects performing four motor imagery tasks (left hand, right hand, both feet, and tongue). The dataset’s label structure necessitates a multi-label classification approach, with four target labels to capture distinct motor imagery tasks. The proposed clustering-based, multi-label classification approach achieves an accuracy of 92% over the all subjects, showing substantial improvement over baseline models on the same dataset. These results confirm that preserving manifold structure in EEG data enables more sensitive and precise clustering and classification, advancing EEG analysis for BCI and neuroprosthetic applications.