Graph neural networks in EEG spike detection.

Area of Science:

• Neuroscience
• Machine Learning
• Medical Technology

Background:

• Interictal epileptiform discharges (IEDs) are crucial biomarkers for epilepsy diagnosis.
• Accurate detection of IEDs in scalp electroencephalography (EEG) remains a challenge.
• Advanced machine learning techniques are needed to improve IED detection.

Purpose of the Study:

• To develop and evaluate novel machine learning architectures for enhanced IED detection in scalp EEG.
• To compare the performance of proposed models against existing methods using the average precision (AP) metric.
• To investigate the efficacy of graph neural networks (GNNs) in identifying epileptogenic activity.

Main Methods:

• Development of two novel GNN-based architectures: FC-GNN and CA-GNN.
• FC-GNN utilizes functional connectivity (FC) maps across frequency sub-bands.
• CA-GNN incorporates an attention mechanism on a complete graph to bypass feature extraction.

Main Results:

• Both FC-GNN and CA-GNN demonstrated superior performance in IED detection on two independent hospital datasets.
• CA-GNN achieved the highest average precision (AP) scores, reaching 0.9788±0.0125 on the Baptist Hospital dataset and 0.9879 on the Temple University Hospital dataset.
• The proposed GNN models significantly outperformed traditional methods like Vanilla Self-Attention and VGG.

Conclusions:

• Graph neural networks (GNNs) show significant promise for the accurate detection of epileptogenic activity.
• The developed FC-GNN and CA-GNN architectures represent a substantial advancement in IED detection.
• This research paves the way for improved 3D source localization in epilepsy patients.