Harmful Brain Activity Classification

Multi-class EEG Classification using Deep Learning
🔗 Kaggle Competition
📌 Developed as part of an official university team project (Team GM2)

🧠 Overview

This project aims to classify harmful brain activity patterns (Seizure, LPD, GPD, LRDA, GRDA) using EEG data. We employed scalogram-based visual representations and trained multiple deep learning models to solve a multi-class classification problem.

👥 Team Members

• Team GM2 (5 members)
• My role:
  • Implemented ResNet18 for EEG classification
  • Proposed and applied Grad-CAM for explainability
  • Also tested a base 2D CNN, which was discarded due to low performance

📂 Dataset

• Source: Kaggle - HMS Harmful Brain Activity Classification
• EEG segments: 50 seconds each
• Labels: Five harmful activity classes (excluding “Others”)

⚙️ Preprocessing

• Montage transformation of raw EEG signals
• Applied bandpass filtering for better results
• Continuous Wavelet Transform (CWT) to generate scalograms
• Resized to 224x224 for CNN input

🧪 Models

• CNN baseline
• ResNet18 (my part)
• EfficientNet
• Vision Transformer (ViT)
• Ensemble of the above three

📈 Performance

Model	Accuracy
ResNet18	~75%
Ensemble	~81%

🛠 Tech Stack

• Python, PyTorch, scikit-learn, Matplotlib, Seaborn

📷 Grad-CAM Visualization