Enhancing Epileptic Seizure Detection with Random Input Selection in Graph-Wave Networks.
Journal:
Annual International Conference of the IEEE Engineering in Medicine and Biology Society. IEEE Engineering in Medicine and Biology Society. Annual International Conference
PMID:
40039673
Abstract
Graph neural networks show strong capability of learning spatial relationships between channels. In recent studies, they greatly advanced automatic epileptic seizures detection via multi-channels scalp electroencephalography (EEG). In this work, we used Graph WaveNet to extract spatial and temporal dependencies of epileptic seizures. However, EEG signals often contain strong noise, leading to unsatisfactory model performance. This study compared effects of four input preprocessing strategies on model robustness. The fast Fourier transform (FFT) features, the input of the network, were preprocessed by intact, hard, learnable, and random selection. Results show that the Graph WaveNet model with random selection (30% dropout) of input FFT features outperforms other benchmarks with an AUROC of 88.57% to detect seizures. Random input selection effectively mitigates over-fitting to noise and promotes the identification of task-related frequencies through global exploration. This preprocessing strategy proves to be a simple yet effective method to improve model robustness, without prior knowledge and additional computational expense.