Source-Space Toolbox for Speech EEG ↗
2024 -A stand-alone pipeline that extends the sensor-space PhD analyses into cortical source space: source estimation, multivariate decoding, multi-atlas parcellation and directed connectivity, built on MNE-Python and scikit-learn.
- Source estimation: projects 64-channel EEG onto fsaverage cortex via a shared forward model (ico-5, 3-layer BEM) with dSPM minimum-norm and LCMV beamformer inverses, per-subject data-rank estimation, and low-memory streaming so whole cohorts run on one machine.
- Decoding: sliding-window classification of ROI source activity with linear SVM, shrinkage LDA and elastic-net logistic regression; nested-CV tuning, vertex/PCA feature modes, pseudo-trial averaging, and spatial leakage correction (symmetric orthogonalisation, regression-based vertex correction).
- Parcellation: 16 speech-network ROIs defined on three atlases (Desikan-Killiany, Schaefer-200, HCP-MMP 1.0) plus custom functional-localizer ROIs projected from volumetric NIfTI masks to the cortical surface.
- Granger causality: parametric (Morf MVAR with Geweke spectral GC; pairwise, block and conditional), nonparametric (Wilson spectral factorisation) and state-space (DARE, MVGC-equivalent) estimators, validated against BSMART to machine precision, with time-reversed GC controls and model-order and window diagnostics.
- Statistics and reproducibility: cluster permutation and TFCE for decoding, permutation task-vs-baseline tests with FDR for connectivity, two-stage subject-parallel and ROI-parallel runners, and theory chapters and technical reports documenting every methodological choice.




