🧠 Is Your Brain Overloaded? — EEG-Based Mental Workload Classifier

A machine learning pipeline that detects binary mental workload states (low vs. high) from EEG signals, achieving 82.7% accuracy with a Support Vector Machine using only 15 carefully selected spectral features.

---

📌 Overview

Mental workload (MWL) is a measure of how much cognitive effort a task demands. Identifying when someone is cognitively overloaded — in real time — has applications in aviation, healthcare, education, and human-computer interaction.

This project builds an end-to-end EEG classification pipeline:

• Raw EEG → Preprocessing → Feature Extraction → Feature Selection → Classification

It uses Power Spectral Density (PSD) features across four EEG frequency bands and a cross-method feature selection strategy (NCA + Random Forest) to identify the 15 most discriminative features before classification.

---

📊 Results

Feature Set	Model	CV Accuracy	Cohen's Kappa
15 Consensus Features	SVM (RBF)	82.7%	0.65
15 Consensus Features	Random Forest	79.5%	0.59
15 Consensus Features	XGBoost	79.6%	0.59
24 NCA Features	SVM (RBF)	80.4%	0.61
24 NCA Features	Random Forest	78.3%	0.57
24 NCA Features	XGBoost	77.3%	0.55

Evaluated under stratified 5-fold cross-validation on 93 recordings from 47 participants.

---

🗂️ Project Structure

EEG-Based_Mental_Workload_Classifier/
│
├── data/                        # Raw EEG recordings (.txt), one per subject per condition
│   ├── sub01_hi.txt
│   ├── sub01_lo.txt
│   └── ...
│
├── preprocessed_data/           # Cleaned EEG arrays saved as .npy files
│   ├── sub01_hi.npy
│   ├── sub01_lo.npy
│   └── ...
│
├── features.csv                 # Extracted PSD features (93 × 58 including subject & workload)
├── ratings.txt                  # Participant subjective workload ratings (1–9 scale)
│
├── exploratory.ipynb            # Data exploration and signal visualisation
├── preprocessing.ipynb          # Full preprocessing pipeline (filter → ASR → re-reference)
├── feature_extraction.ipynb     # PSD feature extraction with sliding windows
└── classification.ipynb         # Feature selection, regression, and binary classification

---

🔬 Pipeline

1. Exploratory Analysis (exploratory.ipynb)

• Loads all 96 raw .txt recordings (14 channels × 19,200 samples at 128 Hz)
• Visualises raw EEG signals across all 14 channels
• Compares low vs. high workload signals side-by-side for all channels
• Confirms uniform recording duration (150 seconds per recording)

2. Preprocessing (preprocessing.ipynb)

Follows the Makoto preprocessing pipeline using MNE-Python and ASRpy:

Step	Method	Details
High-pass filter	Butterworth (MNE)	1 Hz cutoff — removes DC drift
Notch filter	MNE notch_filter	60 Hz — removes line noise
Artifact removal	ASR (cutoff = 5)	Removes eye blinks, muscle artifacts
Re-referencing	Average reference	Removes global noise

3 of 96 recordings failed during ASR and were excluded → 93 recordings retained.

3. Feature Extraction (feature_extraction.ipynb)

• Sliding window: 512-sample windows (4 s), 128-sample step (1 s)
• PSD: Welch's method (nperseg=256) per window per channel
• Bands: Delta (0.5–4 Hz), Theta (4–8 Hz), Alpha (8–13 Hz), Beta (13–30 Hz)
• Averaging: PSD features are averaged across all windows per recording
• Output: 14 channels × 4 bands = 56 features per recording → saved to features.csv

4. Classification (classification.ipynb)

Feature Selection — Consensus of NCA + Random Forest:

• Both methods run over 5-fold CV on the training set
• Features are ranked by cumulative weight (75% threshold)
• NCA selects 24 features; Random Forest selects its own subset
• The 15-feature intersection is used for final classification

Most discriminative features: Delta and theta band power at frontal (AF3, F7, F3) and parieto-occipital (P7, O1, P8) sites.

Classifiers evaluated: SVM (RBF kernel), Random Forest, XGBoost
Best model: SVM with C=100, RBF kernel, trained on 15 consensus features

---

🛠️ Setup & Installation

Python 3.8+ required.

# Clone the repository
git clone https://github.com/your-username/EEG-Based_Mental_Workload_Classifier.git
cd EEG-Based_Mental_Workload_Classifier

# Install dependencies
pip install numpy pandas scipy matplotlib seaborn scikit-learn xgboost mne asrpy

---

▶️ Usage

Run the notebooks in order:

# 1. Explore the raw data
jupyter notebook exploratory.ipynb

# 2. Preprocess all recordings
jupyter notebook preprocessing.ipynb

# 3. Extract PSD features
jupyter notebook feature_extraction.ipynb

# 4. Run feature selection and classification
jupyter notebook classification.ipynb

Note: Update the file paths in each notebook to point to your local data directory before running.

---

📦 Dependencies

Library	Purpose
numpy, pandas	Data handling
scipy	Welch PSD, signal filtering
mne	EEG preprocessing (filtering, re-referencing)
asrpy	Artifact Subspace Reconstruction
scikit-learn	NCA, StandardScaler, SVM, Random Forest, cross-validation
xgboost	XGBoost classifier
matplotlib, seaborn	Visualisation

---

📁 Dataset

This project uses an open-access EEG dataset from 50 male graduate students at Nanyang Technological University.

• Task: SIMKAP multitasking test (18-minute dual-task paradigm)
• Headset: Emotiv EPOC (14 channels, 128 Hz)
• Electrodes: AF3, F7, F3, FC5, T7, P7, O1, O2, P8, T8, FC6, F4, F8, AF4
• Labels: Participant-rated subjective workload (1–9 scale), binarised at threshold 4

---

📄 Paper

This project accompanies the paper:

"Is Your Brain Overloaded? Detecting Mental Workload from EEG Signals"
Folasewa Abdulsalam

Key findings:

• Binary workload classification is tractable from consumer-grade EEG with PSD features alone
• A compact 15-feature set (delta/theta, frontal/occipital) outperforms a larger 24-feature set
• Three-class classification (low/moderate/high) degrades to near-chance — binary framing is more appropriate for this dataset

---