🧠 EEG Research Project: Cognitive Task Classification in ASD

This repository contains the full pipeline for an EEG-based research project focused on power spectral analysis, machine learning, and neural network classification to distinguish cognitive tasks and ASD profiles.

---

📂 Folder Structure

EEG_Research_Project/
├── figures/              # Visual outputs (e.g., PSD plots, scalp maps)
ml_models/
├── 01_asd_vs_td_classification.ipynb
├── 01_asd_vs_td_classification.py
├── 02_asd_vs_td_task_agnostic_model.ipynb
└── 02_asd_vs_td_task_agnostic_model.py
├── preprocessing/        # EEG preprocessing notebooks
│   └── eeg_preprocessing_pipeline.ipynb
├── scripts/              # Python scripts for preprocessing and feature extraction
│   └── eeg_preprocessing_pipeline.py
├── tables/               # CSVs with extracted features and results
├── README.md             # Project overview and structure
├── requirements.txt      # Dependencies
├── LICENSE               # Usage license
└── .gitignore            # Git exclusions

---

🎯 Project Summary

This project involves:

• Preprocessing EEG data collected via MUSE headset.
• Extracting power spectral features (Delta, Theta, Alpha, Beta, Gamma) using Welch’s method.
• Includes machine learning models (SVM, KNN, RF, and Ensemble) for:
  • Classifying participants (ASD vs TD) using task-specific features
  • Predicting ASD vs TD using task-agnostic EEG input (data from any task)
• Neural network models are under development for both use cases above.

---

🔬 EEG Preprocessing Pipeline

The EEG preprocessing pipeline involves the following steps:

1. Raw EEG Loading – Read multiple CSV files containing RAW_ EEG columns
2. NaN and Inf Handling – Use forward-fill, backward-fill, and mean replacement for missing data
3. Conversion to MNE Format – Create RawArray using MNE with 256 Hz sampling
4. Artifact Handling via ICA – Use Independent Component Analysis to detect and remove muscle artifacts using high-frequency PSD checks
5. Power Spectral Density Calculation – Apply Welch’s method to compute absolute band power across Delta, Theta, Alpha, Beta, and Gamma bands
6. Export to CSV – Save averaged power features per file for downstream ML/NN use

---

🤖 Machine Learning Pipeline (ASD vs TD Classification)

1. Data Loading – Features loaded from preprocessed EEG tables

2. Train-Test Split – Stratified sampling with validation

3. Feature Scaling – Using StandardScaler

4. Modeling Techniques:

   • Support Vector Machine (SVM)
   • K-Nearest Neighbors (KNN)
   • Random Forest (RF)
   • VotingClassifier Ensemble (SVM + RF + KNN)

5. Hyperparameter Tuning:

   • KNN: Optimal k selected through manual accuracy comparison
   • SVM: Tuned kernel (linear, rbf) and regularization parameter C
   • Random Forest: Tuned n_estimators, max_depth, and criterion
   • Ensemble: Implemented soft voting using the best-performing base classifiers

6. Best Feature Identification:

   • Feature importances were calculated using permutation importance
   • Applied across all classifiers (SVM, KNN, RF) to measure accuracy drop when features were shuffled

7. Visualizations:

   • Bar plot of the top 2 most important EEG features (from permutation importance)
   • Bar plot of the least 2 important EEG features (from permutation importance)

8. Evaluation Metrics:

   • Accuracy, Precision, Recall, and F1-score using classification_report
   • Compared across all models to select the best-performing classifier

Notebooks:

• 01_asd_vs_td_classification.ipynb – Classifies ASD vs TD using EEG data
• 02_asd_vs_td_task_agnostic_model.ipynb – Predicts ASD vs TD using task-agnostic EEG input

---

🛠️ Requirements

Install dependencies:

pip install -r requirements.txt

---

🚀 Future Work

• Add two neural network models:
  1. NN Model 1: ASD vs TD classification using task-specific EEG features
  2. NN Model 2: ASD vs TD classification using EEG data from any task (task-agnostic binary model)
• Expand hyperparameter optimization and model comparison
• Visualize time-frequency features and statistical comparisons

---

🔗 References

• MNE-Python Documentation
• Welch’s Method