Covariance-Aware Augmentation for Cross-Session BCI: A Multi-Dataset Evaluation

Read the preprint on Preprints.org

Research question

Synthetic EEG augmentation changes the training data available to a decoder. This study evaluates whether augmentation preserves the covariance structure used by common BCI classifiers and how it affects cross-session transfer.

Study design

Four public motor-imagery EEG datasets span 3 to 60 channels. Repeated balanced calibration resampling compares real-only calibration, amplitude-matched Euclidean covariance perturbations, and geometry-preserving log-Euclidean augmentation. The audit measures covariance fidelity, power and conditioning, and downstream accuracy.

Main findings

• Synthetic augmentation did not consistently improve real-only calibration accuracy.
• Same-class geodesic mixup and empirical tangent Gaussian jitter reduced class-mean covariance distance relative to the amplitude-matched Euclidean control.
• Geodesic mixup met the prespecified five-percentage-point non-inferiority margin at 20 calibration trials per class in three primary datasets. The Euclidean control also met that accuracy margin on several datasets, so non-inferiority alone did not distinguish covariance fidelity.
• Power and conditioning showed a mixed profile. Preserving positive-definite covariance structure alone did not ensure positive transfer.

Companion software

TensorEEG provides the R simulation and covariance tools. The real-data audit uses companion Python protocol scripts. TensorEEG-py is the related Python implementation of simulation and augmentation utilities.

View TensorEEG on GitHub