End-to-end EEG analysis pipeline for binary motor imagery classification on the PhysioNet EEGBCI dataset (Schalk et al. 2004). T1 (imagined left fist) vs T2 (imagined right fist), all 109 subjects, with full preprocessing, classical ML baselines, and deep learning models — all under honest cross-subject validation.
Best result: ShallowConvNet at 0.802 balanced accuracy (cross-subject, GroupKFold, 109 subjects). +0.24 over the best classical baseline.
📊 Interactive dashboard: Use streamlit and run locally [Deployed link is coming soon] 📓 Live notebook (Chunk 6, deep learning): https://www.kaggle.com/code/zidanefatuna/eegnet-for-github-com-z-bros-eeg-motorimagery
Median balanced accuracy across 109 subjects, primary metric throughout:
| Method | Within-subject CV | Cross-subject CV (GroupKFold) |
|---|---|---|
| Random chance | 0.500 | 0.500 |
| Classical CSP + LDA | 0.484 | 0.501 |
| Band power + LinearSVC | 0.574 | 0.555 |
| Lateralization indices + LogReg | 0.580 | 0.560 |
| EEGNet (Lawhern 2018) | 0.570 | 0.758 |
| ShallowConvNet (Schirrmeister 2017) | 0.565 | 0.802 |
Within-subject scores cluster around 0.57 — at ~45 trials per subject, no architecture overcomes the sample-size ceiling. Cross-subject reveals the deep-learning advantage clearly.
EEG-MotorImagery/
├── notebooks/
│ ├── 01_data_exploration.ipynb # PhysioNet dataset structure, single-subject inspection
│ ├── 02_preprocessing.ipynb # Pipeline: filter → bad channels → ICA → AutoReject
│ ├── 03_erp_tfr_analysis.ipynb # ERP, mu/beta ERD, cluster permutation tests
│ ├── 04_classical_ml.ipynb # Lateralization, band power, CSP at single-subject
│ ├── 05_classical_multisubject.ipynb # Same pipeline at population scale (109 subjects)
│ └── 06_deep_learning.ipynb # EEGNet + ShallowConvNet (runs on Kaggle)
├── src/
│ ├── preprocessing.py # Composable preprocessing functions
│ ├── features.py # ERP/TFR utilities + ML feature extraction
│ ├── visualization.py # Plotting helpers
│ ├── stats.py # Cluster permutation tests
│ ├── multisubject.py # Multi-subject orchestration (cache, CV drivers, transformers)
│ ├── eegnetmods.py # EEGNet + ShallowConvNet PyTorch architectures
│ └── torch_data.py # Dataset, training loop, sklearn-compatible wrapper
├── dashboard/ # Streamlit app — interactive results explorer
├── docs/
│ └── methodology.md # Detailed methodology writeup
└── README.md # this file
The methodology is fully documented in
docs/methodology.md. Quick summary:
Preprocessing (Chunk 2) — load EDF, standardize channel names with the standard_1005 montage, notch 60 Hz, bandpass 1–40 Hz, statistical bad-channel detection (interpolate, don't drop), average reference, Picard ICA with 20 components and Fp1 as pseudo-EOG, epoch −0.5 to +4.0 s around cue, baseline correct, AutoReject.
Honest cross-subject validation — GroupKFold by subject_id, balanced accuracy as primary metric (4 subjects had ~1.5× class imbalance after AutoReject), modal sfreq (160 Hz) for population concatenation, interpolated bads instead of dropped channels (keeps feature dimension constant).
Pre-registered predictions — for the deep learning chunk specifically, predictions were registered in §1 of the notebook before any model was trained. Three of four resolved as predicted; the one miss (ShallowConvNet beat EEGNet by 0.044, outside the ±0.03 pre-registered band) produced the chunk's most interesting finding.
The within-subject ceiling is sample-size, not architecture. Classical features, EEGNet (~2.8K params), and ShallowConvNet (~107K params) all converge to ~0.57 within-subject. At ~45 trials per subject and ~7 trials in the inner validation set, early stopping fires on noise. This is a property of the dataset, not the methods.
Cross-subject unlocks deep learning's advantage. With 3,000+ training
trials per fold, both deep architectures cleanly exceed classical baselines.
ShallowConvNet wins because its temporal conv → spatial conv → square → log-pool architecture is the differentiable analog of CSP — and the
log-variance assumption is correct for motor imagery (mu/beta band power
desynchronization is the textbook signature).
Spatial filter analysis reveals why ShallowConvNet wins. Topomap visualization of cross-subject spatial filters shows:
- ShallowConvNet's top filters localize to motor cortex (C3, C1, CP3, CP4, FCz, C4 dominating top-3 channels) — the expected anatomy for hand motor imagery.
- EEGNet's top filters concentrate on frontal-channel artifacts (F8, AF8, F7, FT7, AF7) — replicating the same failure mode classical CSP showed in Chunk 5 §6 (population-averaged CSP filters localized to Fp1, Fp2, F7, F8, T8). Residual EMG/oculomotor activity that survived ICA correlates with T1/T2 labels and the network exploits it.
So Shallow's 4-point advantage over EEGNet is a quality difference, not just a numerical one: it's solving the task using the features the protocol was designed to elicit, while EEGNet is partly exploiting correlated artifact.
Notebooks 01–05 run locally on CPU. Standard environment:
git clone https://github.com/Z-bros/EEG-MotorImagery.git
cd EEG-MotorImagery
pip install -r requirements.txt
jupyter notebook notebooks/01_data_exploration.ipynbNotebook 06 (deep learning) runs on Kaggle with the free T4 GPU.
Reproduction instructions are in the first cell of the notebook. Local
GPU also works if you adapt the /kaggle/working/ and /kaggle/input/
paths.
Dashboard:
cd dashboard
streamlit run app.pyMNE-Python (preprocessing, ERP/TFR, topomap visualization), scikit-learn (classical ML and CV infrastructure), AutoReject (epoch rejection), PyTorch (EEGNet, ShallowConvNet), Streamlit (dashboard).
- Results are good for honest-methodology PhysioNet motor imagery (~0.80 cross-subject is in the literature range of 0.75–0.90) but well below what a deployable BCI needs (~0.95+).
- Papers claiming 95–99% accuracy on this dataset typically use trial-randomized splits without subject grouping (leakage), peak-not-mean accuracy, cherry-picked responder subsets, or one-vs-rest reframings — none of which are methodologically defensible. Our 0.80 is the honest number.
- The artifact-contamination finding for EEGNet is conditional on this dataset and our ICA pipeline. With tighter EMG/oculomotor component rejection, EEGNet might localize cleanly.
MIT License
Copyright (c) 2026 Zidane Fatuna
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
If this work is useful to you, citation appreciated but not required:
Zidane Fatuna (2026). EEG Motor Imagery Classification — Portfolio Project.
GitHub: https://github.com/Z-bros/EEG-MotorImagery
Dataset: Schalk, G., et al. (2004). BCI2000: A General-Purpose Brain-Computer Interface (BCI) System. IEEE TBME 51(6):1034-1043.
- PhysioNet for hosting the EEGBCI dataset
- MNE-Python developers
- Lawhern et al. for EEGNet (2018)
- Schirrmeister et al. for ShallowConvNet (2017)
- Anthropic Claude (Opus 4.7) for code review and methodology discussion during portfolio development