Dynamic Graph Convolutional Network with Dilated Convolution for Epilepsy Seizure Detection.

Area of Science:

• Neuroscience and Biomedical Engineering
• Signal Processing and Machine Learning

Background:

• Electroencephalogram (EEG) is crucial for measuring brain activity and has been widely used for automated epileptic seizure detection.
• Existing methods using Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM), and Graph Convolutional Networks (GCNs) face limitations due to assumptions of Euclidean structure or fixed graph connectivity.

Purpose of the Study:

• To address limitations in current automated seizure detection algorithms.
• To propose a novel Dynamic Graph Convolutional Network with Dilated Convolution (DGDCN) for enhanced EEG-based seizure detection.

Main Methods:

• Developed a Dynamic Graph Convolutional Network with Dilated Convolution (DGDCN) algorithm.
• Employed a spatiotemporal attention mechanism to dynamically construct a task-specific adjacency matrix for GCNs.
• Integrated a dilated convolutional module to expand the receptive field and capture long-range temporal dependencies.

Main Results:

• The DGDCN model achieved Area Under the Curve (AUC) values of 88.7% on 12-second EEG segments and 90.4% on 60-second segments.
• Demonstrated competitive performance compared to current state-of-the-art seizure detection methods on the TUSZ dataset.

Conclusions:

• The proposed DGDCN model effectively captures localized spatial and temporal dependencies and long-range temporal information in EEG signals.
• The dynamic graph construction and dilated convolutions offer a significant advancement in automated epileptic seizure detection accuracy.