Multiclass Emotion Recognition from EEG Signals

Project Overview

This project implements a robust system for multiclass emotion recognition using EEG signals. It utilizes the GAMEEMO dataset to classify four distinct emotional states: Boring, Calm, Scary, and Funny.

The core innovation of this project is a Unique Dual-Branch Parallel Hybrid Approach that fuses the strengths of Machine Learning (Random Forest) and Deep Learning (CNN). This method aims to balance the interpretability of hand-crafted features with the high performance of deep feature learning, achieving a recognition accuracy of approximately 90%.

Features

• Dataset: GAMEEMO (28 subjects, 14 channels, 128Hz sampling rate).
• Preprocessing Pipeline:
  • Artifact removal and channel normalization.
  • Bandpass filtering (1-45 Hz) and notch filtering (50/60 Hz).
  • Epoch cleaning (handling NaNs and flat channels).
• Feature Extraction:
  • Power Spectral Density (PSD) analysis.
  • Band-power feature extraction across Delta, Theta, Alpha, Beta, and Gamma bands.
• Modeling:
  • Random Forest (RF): Trained on extracted spectral features for interpretability.
  • Convolutional Neural Network (CNN): Trained on raw EEG signals for complex pattern recognition.
  • Hybrid Fusion: A weighted voting mechanism combining RF and CNN predictions.
• Analysis:
  • Feature importance ranking.
  • Band-specific accuracy analysis.

Prerequisites

Ensure you have the following Python libraries installed:

pip install mne scipy matplotlib numpy pandas scikit-learn kagglehub
# + Deep Learning framework (TensorFlow/Keras or PyTorch depending on environment)

Structure

• NDS_PRJ_cleaned.ipynb: The main Jupyter Notebook containing the end-to-end pipeline: data loading, preprocessing, feature extraction, model training, and evaluation.
• nds (5).pdf: Project proposal and requirements document.
• fix_notebook.py: Utility script for notebook maintenance.

Implementation Details

1. Data Acquisition: The project automatically downloads the GAMEEMO dataset using kagglehub.
2. Processing: EEG data is converted to MNE RawArray objects and preprocessed to remove noise.
3. Classification:
   • Branch 1 (ML): Extracts 70 spectral features (14 channels × 5 bands) and trains a Random Forest classifier (n_estimators=300).
   • Branch 2 (DL): Processes raw signal segments using a CNN.
   • Fusion: The final prediction is a weighted average of the probabilities from both models.
4. Results: The fusion model typically outperforms individual models, with the best observed weight distribution around 0.8 (RF) / 0.2 (CNN).

Usage

1. Open NDS_PRJ_cleaned.ipynb in Jupyter Notebook or Google Colab.
2. Run the cells sequentially.
3. The notebook will:
   • Download the dataset.
   • Preprocess the EEG data.
   • Train the models.
   • Display accuracy metrics, confusion matrices, and feature importance plots.