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

Source abstract

Covariance augmentation expands small calibration sets for brain–computer interfaces (BCIs), but valid synthetic matrices need not preserve real-trial variation or improve cross-session decoding. We examined the relationship between covariance construction, distribution fidelity, and decoding performance. We compared real-only training with ambient noise followed by positive-definite projection, isotropic and class-dependent log-space noise, and same-class log-Euclidean interpolation on four motor-imagery electroencephalography (EEG) datasets with 3–60 channels. Using calibration draws paired across methods, we evaluated 83 subjects at 20 trials per class. Augmentation did not consistently improve balanced accuracy. Log-Euclidean interpolation changed accuracy by -0.25 to -0.01 percentage points (pp) and passed the additional Holm-corrected non-inferiority tests at a -5 pp group-mean margin. To determine whether near-neutral accuracy also reflected preserved variation, we compared synthetic samples with source trials excluded from calibration. Interpolation retained 62%–65% of the reference within-class scatter and had a larger energy distance than repeated real trials, despite remaining close to calibration class centers. Matched arithmetic interpolation lost more accuracy on two datasets; the methods differed little on the other two. We next tested the role of the classifier's covariance estimator. On BNCI2014_001 at 30 trials per class and augmentation ratio 3, changing fixed-ridge linear discriminant analysis (LDA) to automatic-shrinkage LDA shifted an isotropic-noise effect from -7.35 to +0.10 pp against the respective real-only baselines. Log-Euclidean interpolation and repeated real trials both lost approximately 1.9 pp under shrinkage. Finally, synthetic experiments with preserved labels and matched augmentation counts yielded similar gains from ambient and log-space noise under controlled covariance rotations. These results show that augmentation outcomes depend on both the generated distribution and classifier estimation. Positive definiteness provides valid covariance inputs; assessing fidelity and decoding utility requires independent distribution references and matched classifier comparisons.