EEG-based Harmful Brain Activity Classification using Deep Learning

View Project Repository

🍀 Overview

This study explores how deep learning models can detect neurological disorders from EEG data, while addressing interpretability challenges using model explainability tools.

---

🍀 Dataset

• The project is based on the Kaggle competition HMS - Harmful Brain Activity Classification.
• We used 50-second EEG segments labeled as one of five harmful brain activity types (excluding “Others”).
  

---

🍀 Process

• Preprocessed raw EEG using montage transformation and bandpass filtering
• Converted EEG into scalogram images using Continuous Wavelet Transform (CWT)
• Trained and compared multiple models including:
  • ResNet18 (my implementation)
  • EfficientNet
  • Vision Transformer (ViT)
• Achieved ~81% accuracy using ensemble, ~75% with ResNet18

---

🍀 My Contribution

• Implemented and trained ResNet18 for EEG classification
• Applied Grad-CAM for model interpretability
• Tested baseline 2D CNN → discontinued due to low generalization
• Collaborated on model evaluation and ablation studies
  

---

🍀 Limitations

• Grad-CAM does not support ViT well → heatmaps were not interpretable
• Future work could include SHAP, Attention Rollout, or LRP for ViT explainability