A Multi-Dataset Comparative Analysis of Brain-Computer Interface Classification Techniques Using Public Neural Data

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Brain-Computer Interface (BCI) systems allow direct communication between the human brain and external devices through the analysis of electroencephalography (EEG) signals; however, the performance and generalization capability of EEG classification models are highly dependent on characteristics of the dataset, subject variability, and evaluation parameters. In this research article, we demonstrated a comparative benchmarking framework and evaluated machine learning-based EEG classification across three publicly available datasets: BCI Competition IV Dataset 2a, PhysioNet-EEG Motor Movement/Imagery, and an Open-Closed Eyes EEG dataset. To ensure fair and reproducible evaluation, a unified processing pipeline comprising EEG preprocessing, Common Spatial Pattern (CSP) feature extraction, and five machine learning classifiers, specifically Linear Discriminant Analysis (LDA), Support Vector Machine (SVM), Random Forest (RF), Multi-Layer Perceptron (MLP), and Extreme Gradient Boosting (XGBoost), was presented in this paper. The experimental results show substantial variation in performance across datasets. The RF achieved the highest five-fold cross-validation accuracy on the BCI Competition IV Dataset 2a, which is 55.36 ± 1.54%, while the SVM gave the best subject-independent Leave-One-Subject-Out (LOSO) accuracy of 37.50 ± 8.09%. On the other hand, with the PhysioNet dataset, the LDA achieved the highest accuracy of 60.23 ± 3.01%. In the case of the Open-Closed Eyes dataset, XGBoost achieved an accuracy of 79.82 ± 5.25%. This research demonstrated that EEG classification performance is strongly influenced by dataset complexity and inter-subject variability, highlighting the importance of standardized evaluation protocols and robust benchmarking frameworks for EEG-based BCI research.