EEG Seizure Detection with Graph-Based and Deep Learning Models

Machine Learning

Deep Learning

Graph Neural Networks

EEG Analysis

Healthcare AI

Transformers

Neuroinformatics

École Polytechnique Fédérale de Lausanne (EPFL) • February – June 2025

Machine Learning Researcher

Project Overview

Completed as a semester project in the Network Machine Learning course at EPFL, this project investigated multiple neural architectures—including EEGNet, EEGNet+Transformer, GCN+BiLSTM, and NeuroGNN—for detecting epileptic seizures from EEG recordings. The dataset included over 75 patients from the Temple University Seizure Corpus. In a team of four, we compared model accuracy, recall, and robustness on both public and private test sets, revealing key trade-offs between performance and generalizability.

Challenges

•Handling class imbalance in seizure vs. non-seizure EEG recordings
•Evaluating trade-offs between model complexity and clinical robustness
•Designing fair comparisons between spatial (GNN-based) and temporal (CNN/RNN-based) approaches
•Hyperparameter optimization and architecture tuning for EEGNet+Transformer
•Balancing interpretability with prediction accuracy in a sensitive medical context

Key Achievements

•Tested 3 advanced models on real-world EEG seizure datasets, achieving up to 90% validation accuracy
•EEGNet+Transformer showed best generalization (81.8% private test accuracy) and emerged as most clinically viable
•Implemented and extended NeuroGNN with attention-based edge construction and dynamic graph updates
•Performed hyperparameter sweeps and attention module testing (ECA), yielding detailed performance insights
•Produced a scientific report comparing results across multiple performance metrics and clinical criteria
•Achieve the 2nd place in the EPFL Kaggle Leaderboard

Technologies Used

Python

PyTorch

Graph Neural Networks (GNNs)

EEGNet

Transformers

Bidirectional LSTM