EEG-based Harmful Brain Activity Classification using Deep Learning

View Project Repository

Why This Matters

EEG is central to diagnosing neurological conditions such as seizures and brain injury, but reading it is slow, costly, and subjective — expert agreement between clinicians is often low, and distinguishing visually similar abnormal patterns is error-prone. An accurate and interpretable automated system could give clinicians an objective second opinion they can actually trust. This project builds a model to classify five representative abnormal-EEG patterns — Seizure, LPD, GPD, LRDA, and GRDA — and, just as importantly, to show why it makes each prediction.

Overview

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

Dataset

• 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 (0.5–40 Hz)
• 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 a weighted ensemble, and ~75% with ResNet18 alone

My Contribution

• Implemented and trained ResNet18 for EEG classification (baseline accuracy ~75%, a ~30%p gain over a plain 2D CNN)
• Applied Grad-CAM for model interpretability
• Designed and tested a baseline 2D CNN → discontinued due to low generalization
• Collaborated on model evaluation and ablation studies

Results & Error Analysis

• Strong sensitivity on LPD and GRDA classes, capturing distinct EEG patterns within the scalograms
• Some LRDA samples were initially misclassified as Seizure or LPD; this was mitigated through ensemble learning, lifting final system accuracy to ~81%
• Grad-CAM confirmed the model attends to clinically meaningful EEG signals rather than background noise — supporting the transparency needed for AI-assisted diagnosis

Limitations

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