EEG Alpha Waves – Codabench Challenge (Rodrigues2017)

Group members

• Alexandre HEYMANN
• Mohamed MOHAMED EL BECHIR
• Noé AMAR
• Olivier Fontaine
• Alexandre Zénou

Overview

This repository defines a fully working Codabench challenge for a binary classification task based on EEG recordings.

It includes:

• A complete competition bundle (bundle.zip)
• Ingestion and scoring programs
• Prepared dataset splits (train / public test / private test)
• Competition web pages (pages/)
• A mandatory starting kit notebook (starting_kit.ipynb)
• A reproducible data preparation script (tools/setup_data.py)

---

Dataset and Data Origin

We use the EEG Alpha Waves dataset, available in MOABB under the name Rodrigues2017.

Original dataset source (Zenodo): https://zenodo.org/record/2348892

The dataset is automatically downloaded via MOABB/MNE when running:

python tools/setup_data.py

---

Task

Goal

Build a binary classifier from EEG-derived features.

Feature Extraction

During dataset construction:

• EEG signals are segmented into 2-second windows.
• For each channel and window, we compute:
  • Mean
  • Standard deviation
  • Alpha band power (8–12 Hz) using Welch PSD integration.

This produces a tabular dataset compatible with standard scikit-learn models.

Labels

Binary labels are constructed:

• From event annotations when available.
• Otherwise, using a fallback time-based split (early vs late recording).

---

Evaluation Strategy

Metric

The challenge uses Accuracy, defined as:

$$\text{Accuracy} = \frac{\text{Number of correct predictions}}{\text{Total number of predictions}}$$

Since the dataset is balanced, accuracy is appropriate.

The metric is implemented in:

scoring_program/scoring.py

Splitting Strategy

We use a subject-wise split:

• Train
• Public test
• Private test

This prevents data leakage across subjects and evaluates generalization to unseen participants.

---

Submission Format

Participants must submit a single Python file:

submission.py

Containing:

def get_model():
    return ...

The ingestion program will call:

model.fit(X_train, y_train)
model.predict(X_test)

---

Local Usage

1. Install dependencies

pip install -r requirements.txt

2. Generate dataset splits

python tools/setup_data.py

3. Test ingestion

python ingestion_program/ingestion.py \
  --data-dir dev_phase/input_data/ \
  --output-dir ingestion_res/ \
  --submission-dir solution/

4. Test scoring

python scoring_program/scoring.py \
  --reference-dir dev_phase/reference_data/ \
  --prediction-dir ingestion_res/ \
  --output-dir scoring_res/

5. Create bundle

python tools/create_bundle.py

This generates:

bundle.zip

---

Deliverables

Final submission includes:

• GitHub repository link
• Codabench competition URL
• Working bundle
• Starting kit notebook