Enhancing Epileptic Seizure Detection with Random Input Selection in Graph-Wave Networks.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Graph neural networks (GNNs) excel at analyzing spatial relationships in multi-channel electroencephalography (EEG) data for seizure detection.
• Existing GNN models for automatic epileptic seizure detection struggle with performance degradation due to noise in EEG signals.

Purpose of the Study:

• To investigate the impact of different input preprocessing strategies on the robustness of Graph WaveNet for epileptic seizure detection.
• To enhance the performance of GNNs in noisy EEG environments.

Main Methods:

• Utilized Graph WaveNet to model spatial and temporal dependencies in EEG data.
• Compared four preprocessing strategies for fast Fourier transform (FFT) features: intact, hard selection, learnable selection, and random selection (30% dropout).

Main Results:

• The Graph WaveNet model employing random selection of input FFT features achieved the highest performance.
• This approach resulted in an Area Under the Receiver Operating Characteristic Curve (AUROC) of 88.57% for seizure detection.
• Random selection effectively reduced over-fitting to noise and identified task-relevant frequencies.

Conclusions:

• Random input feature selection is a simple, effective, and computationally inexpensive method to enhance GNN robustness for noisy EEG seizure detection.
• This strategy improves model generalization without requiring prior domain knowledge.
• The findings offer a practical approach to improving automated seizure detection systems.