mi_riemann_benchmark.py runs the @phoebe/signal Riemannian pipeline's reference
algorithm (pyRiemann Covariances → MDM) against a CSP+LDA baseline on a real
public motor-imagery dataset (PhysioNet EEGBCI, subject 1, left-vs-right-hand
imagery, MNE-fetched). Requires pip install mne pyriemann scikit-learn.
Result (subject 1, 45 epochs, 64 ch, 5-fold CV):
- Riemannian MDM: 68.9%
- CSP + LDA: 64.4%
- Riemannian MDM reproduces published EEGBCI accuracy and beats CSP+LDA — the 181.4.2.2 acceptance.
The @phoebe/signal TypeScript riemannianDistance is separately validated to match
pyRiemann's distance_riemann on real MI-derived covariance matrices — see
src/riemann-validation.spec.ts (reference distance 0.7684150 on the embedded
EEGBCI covariance fixtures).