A Standardized EEG Data Acquisition and Experimental Setup

Project Overview

This project outlines a rigorous protocol for generating a standardized EEG dataset. The resulting data is engineered for high-performance applications in computational neuroscience and machine learning, specifically targeting:

• Brain-Computer Interfaces (BCI)
• Deep Learning Research (CNNs, LSTMs, Transformers)

The setup follows global neuroimaging standards to ensure the dataset is interoperable, reproducible, and ready for public research distribution.

---

Objectives

• Scalable Design: Create a clean, modular EEG experimental pipeline.
• Multi-Subject Collection: Acquire diverse data across a broad subject pool.
• Standardization: Adhere to BIDS and IFCN guidelines for seamless data sharing.
• AI-Ready: Provide structured, well-labeled data for immediate ML/DL model training.

---

Experimental Standards

This protocol adheres to the following international frameworks:

Standard	Description	Reference
IFCN Guidelines	Clinical and research EEG recording standards	Clinical Neurophysiology
EEG-BIDS	Brain Imaging Data Structure for EEG	BIDS Specification
ERP Design	Principles for Event-Related Potential techniques	Luck (2014)

---

Hardware & Acquisition

Device Compatibility

The protocol supports research-grade and "prosumer" mobile EEG devices:

• OpenBCI (Cyton/Daisy boards)
• Emotiv (EPOC+/Flex)

Electrode Placement

We utilize the International 10–20 System to ensure consistent spatial mapping across subjects.

Signal Parameters

To maintain high signal fidelity, the following parameters are enforced:

• Sampling Rate: 250–500 Hz
• Resolution: $\ge$ 16-bit
• Electrode Impedance: $< 5$ k$\Omega$
• Hardware Filter: 50 Hz Notch filter (Power-line noise removal)

---

Subject Preparation & Ethics

Inclusion Criteria

• Age: 18–40 years.
• Health: Healthy individuals with no history of neurological or psychiatric disorders.

Pre-Experiment Protocol

1. Caffeine: Avoid intake 2–3 hours prior to the session.
2. Rest: Minimum 7–8 hours of sleep the previous night.
3. Physical: Removal of all metallic jewelry/accessories.

Ethical Compliance

• Full informed consent obtained from all participants.
• Data is strictly anonymized using unique Subject IDs (e.g., sub-01).
• Reference: WHO Research Ethics

---

Experimental Protocol

Each session is divided into three distinct phases:

Phase 1: Baseline Recording

Establishment of "Resting State" EEG activity (Eyes Open/Eyes Closed).

Phase 2: Stimulus / Task Phase

• Stimulus Duration: 30–60 seconds.
• Rest Interval: 10–15 seconds.
• Trials: 10–30 repetitions per condition.
• Tasks: Musical stimuli, visual tracking, cognitive problem-solving.

Phase 3: Self-Assessment Labeling

Post-trial subjective feedback including:

• Emotion: Valence (Pleasure) and Arousal (Excitement).
• Workload: Task difficulty and attention levels.

---

Dataset Structure

The data follows the BIDS-inspired directory hierarchy:

dataset/
 ├── sub-01/
 │   ├── eeg/
 │   │   ├── raw_eeg.csv     # Time-series voltage data
 │   │   ├── events.csv      # Stimulus markers and timestamps
 │   ├── labels.json         # Self-assessment scores
 │   └── metadata.json       # Device and session info
 ├── sub-02/
...

---

Metadata Example

{
  "sampling_rate": 250,
  "channels": ["F3", "F4", "T7", "T8"],
  "experiment_type": "emotion",
  "stimulus": "music"
}

---

Signal Processing Pipeline

Preprocessing

• Band-pass filter: 0.5–45 Hz
• Notch filter: 50 Hz
• Artifact removal using ICA

Reference:
https://mne.tools/stable/index.html

---

Feature Extraction

Frequency Bands

• Delta (0.5–4 Hz)
• Theta (4–8 Hz)
• Alpha (8–13 Hz)
• Beta (13–30 Hz)
• Gamma (>30 Hz)

---

Extracted Features

• Power Spectral Density (PSD)
• Differential Entropy
• Hjorth Parameters
• Wavelet Transform

---

Machine Learning Models

Recommended Models

• Support Vector Machine (SVM)
• Random Forest
• Convolutional Neural Networks (CNN)
• Long Short-Term Memory (LSTM)

---

Reference Models

• EEGNet (Compact CNN for EEG)
  https://arxiv.org/abs/1611.08024

• Deep Learning Review
  https://www.sciencedirect.com/science/article/pii/S0957417419302743

---

Benchmark Datasets (Reference)

DEAP Dataset

https://www.eecs.qmul.ac.uk/mmv/datasets/deap/
Paper: https://ieeexplore.ieee.org/document/5871728

---

SEED Dataset

http://bcmi.sjtu.edu.cn/~seed/
Paper: https://ieeexplore.ieee.org/document/7104132

---

PhysioNet EEG Dataset

https://physionet.org/content/eegmmidb/1.0.0/

---

TUH EEG Corpus

https://www.isip.piconepress.com/projects/tuh_eeg/html/downloads.shtml

---

Dataset Publication

Recommended platforms:

• Kaggle
  https://www.kaggle.com/datasets

• Zenodo (DOI generation)
  https://zenodo.org/

• PhysioNet
  https://physionet.org/

---

Best Practices

• Maintain consistent experimental conditions
• Minimize noise and movement artifacts
• Ensure accurate and consistent labeling
• Validate data quality before publishing
• Provide baseline ML model results

---

Limitations

• EEG signals are highly sensitive to noise
• Subject variability affects consistency
• Limited spatial resolution compared to fMRI

---

License

Specify dataset license before publishing.

Recommended options:

• MIT License
• CC BY 4.0

---