Benchmarking Classification Pipelines Within and Across Sessions on the PhysioNet EEG Dataset
Date:
Authors: Yiming Shen, David Degras (University of Massachusetts Boston)
Abstract
Electroencephalography (EEG) remains a cornerstone of noninvasive brain-computer interfaces (BCI) due to its affordability and temporal resolution. However, classification accuracy often drops between sessions due to non-stationarity caused by changes in electrode placement, hardware, or physiology.
This talk presents a rigorous benchmark of feature extraction and classification methods on the open PhysioNet MI dataset (109 participants, 12 sessions, 64-channel EEG), aiming to replicate results from the Mother of all BCI Benchmark (MOABB) and explore new strategies for cross-session classification.
Methodology
We utilized nested cross-validation to select hyperparameters and evaluate the following pipelines:
- Feature Extraction: Log-variance, Common Spatial Patterns (CSP), Filter Bank CSP (FBCSP), Covariance-Matrix Tangent Space (TS), and Augmented Covariance Matrices mapped to Tangent Space (ACM-TS).
- Classifiers: Linear Discriminant Analysis (LDA), Support Vector Machines (SVM), Elastic Net Logistic Regression (EL), Riemannian Minimum Distance to Mean (MDM), and Multinomial Logistic Regression (LR).
- Cross-Session / Domain Adaptation (DA): To address session drift, we employed Transfer Component Analysis (TCA), Subspace Alignment (SA), and Correlation Alignment (CORAL), utilizing strategies for DA with 3 or more sessions.
Results & Contributions
- Performance: Mean accuracies across subjects and sessions ranged from 54% to 72%, significantly exceeding the chance level of 33%.
- Reproducibility: By combining a modular feature-classifier design with rigorous validation, this work provides a realistic benchmark for multi-class MI decoding.
- Open Source: We introduced an open-source R implementation to allow for rapid prototyping and fair comparison of future methods.