Projects / private repository

NeuroForge: cross-subject EEG motor-imagery decoding

independent

A reproducible brain-computer-interface pipeline that measures how much motor-imagery decoding accuracy drops when a model meets a new person.

Why it matters

Decoders that work for one person often fail on the next. Measuring that cross-subject drop is the first step toward BCIs that generalize.

Stack

Python · MNE-Python · MOABB · scikit-learn · pyRiemann · EEGNet · pytest · GitHub Actions

• eeg
• bci
• machine-learning
• python
• reproducibility

What it does

• A 12-week curriculum and codebase covering EEG simulation, filtering, and decoding with band-power features, common spatial patterns, and Riemannian classifiers.
• A reproducible pipeline on the MOABB framework that decodes motor-imagery EEG from public datasets and measures the accuracy drop from within-session to leave-one-subject-out evaluation, quantifying cross-subject domain shift.
• Compares classical pipelines (CSP + LDA; Riemannian covariance on the SPD manifold) with a deep model (EEGNet) using Euclidean Alignment, reproducing Eder et al. (2024): the deep nets showed no significant cross-subject advantage over Riemannian methods.
• An automated test suite and continuous integration keep development reproducible.