A self-supervised graph network with time-varying functional connectivity for seizure prediction

Seizure prediction based on scalp EEG can improve the quality of life for patients with epilepsy. It is challenging to implement seizure prediction methods, limited to the heterogeneity of patients and the irregular spatial, temporal and spectral evolution of epileptic EEG. To tackle these, we propose a self-supervised graph representation learning model captures the spatial–temporal-spectral responses with limited labeled data. Specifically, we propose a method to infer the time-varying functional connectivity corresponding to the brain network. Meanwhile, we design a temporal-spectral analysis network with wavelet-enhanced and a spatiotemporal analysis network via graph convolution to analyze the brain network across multiple spatial, temporal, and spectral scales. Finally, we utilize a twin-structured contrastive learning framework to learn the intrinsic representations of epileptic EEG and pre-train the proposed model with unlabeled data. Our method is evaluated on the CHB-MIT and Siena dataset. On CHB-MIT dataset, the method achieves an AUC of 0.991, an accuracy of 0.990, a sensitivity of 0.992, and a false positive rate of 0.012. And the method also shows commendable performance on Siena dataset, underscoring its robustness and reliability across diverse datasets. Furthermore, it still demonstrate competitive performance when the amount of labeled data is reduced. The results indicate our model has the potential for predicting seizure in practical application. They also provide a new perspective for analyzing the multidimensional, multi-scale brain network.