Abstract
In this paper, we propose a personalized seizure detection and classification framework that quickly adapts to a specific patient from limited seizure samples. We achieve this by combining two novel paradigms: graph neural networks (GNN) and meta-learning. We train a Meta-GNN based classifier that learns a global model from a set of training patients such that this global model can eventually be adapted to a new unseen patient using very limited samples. We apply our approach on the TUSZ-dataset, one of the largest and publicly available benchmark datasets for epilepsy. We show that our method outperforms the baselines by reaching 82.7% on accuracy and 82.08% on F1 score after only 20 iterations on new unseen patients.
Introduction
Epilepsy affects over 60 million people worldwide, necessitating constant EEG monitoring. Traditionally, this process is performed manually by experts—a task that is slow, expensive, and prone to fatigue. While deep learning (specifically CNNs) has automated some of this, it faces two significant hurdles: it often ignores the physical and functional relationships between electrodes (geometry), and it struggles with high inter-patient variability. A model trained on a large population might fail on a new individual due to unique physiological differences. This study introduces a system that is both spatially aware (using graphs) and rapidly adaptable (using meta-learning) to handle these unseen patients with minimal data.
Methods
The researchers developed a pipeline merging spatial modeling with few-shot adaptation:
- EEG as a Graph: The brain's electrode network is represented as a graph where nodes are electrodes and edge weights are calculated using a Gaussian kernel based on physical distance. Node features are extracted using Fourier coefficients.
- GNN Architectures: Two types were evaluated: Graph Convolutional Networks (GCN) with a fixed structure, and Graph Attention Networks (GAT) which dynamically learn electrode connections.
- Meta-Learning (MAML): Model-Agnostic Meta-Learning is used to train the model to "learn how to learn" a new patient's patterns quickly.
- Modified Loss Function: To prevent memorization overfitting, a regularization term was added to the MAML "outer loop," forcing the model to rely more on the specific "support set" of the new patient.
Experimental Setup
The study utilized the Temple University Hospital EEG Seizure Corpus (TUSZ), one of the largest public datasets for adult seizures:
- Training: 97 patients (filtered for those with >4 seizures).
- Testing: 26 "unseen" patients (filtered for those with >2 seizures).
- Preprocessing: EEG signals were split into 10-second windows and normalized.
Key Results and Insights
The Meta-GNN approach (specifically GAT-ML) showed impressive gains over standard baselines.
- Efficiency in the Clinic: The model achieves high accuracy after only 20 iterations of fine-tuning, hitting 82.70% accuracy. In a hospital setting, this means a system could become personalized within minutes of a patient being hooked up to an EEG.
- Personalization is Mandatory: The "Global GAT" model (trained on everyone but not fine-tuned) performed significantly worse, confirming that a universal model is insufficient for clinical-grade EEG analysis.
- GAT vs. GCN: While both performed well at detection, GAT was superior for classification (focal vs. generalized seizures). This suggests that dynamic functional connectivity is more informative than static physical distance for distinguishing seizure types.
Conclusion
This framework demonstrates that combining GNNs with meta-learning allows for rapid personalization in seizure detection. Future work will explore different window sizes, a broader spectrum of seizure types, and entirely learned graph structures that do not rely on fixed electrode placement.
Summary updated