EEG-TGC: A Novel Self-Supervised Method based on Temporal-Graph Contrast for Seizure Detection

Epilepsy is a complex neurological disorder that can be diagnosed using electroencephalography (EEG). Although high performance achieved in previous studies on seizure detection, several challenges persist. Firstly, while EEG recordings are typically easily available, the annotation of these records imposes a significant burden on clinicians. Most of the unlabeled data cannot be used directly, leading to wastage. Secondly, EEG signals are commonly recorded from multiple electrodes; however, some studies have overlooked the critical spatial distribution information among these electrodes, leading to suboptimal classification performance. In this study, we address these challenges by proposing a novel self-supervised learning method based on temporal graph contrast, named EEG-TGC. It effectively utilizes a large volume of unlabeled data. By introducing node and graph contrast, our method adeptly captures the robust spatial topological information of EEG graphs. Our approach is evaluated on a large public EEG dataset, TUSZ. Experimental results demonstrate that our method achieves performance comparable to supervised learning using 100% labeled data, even with only 10% labeled data.Clinical relevance- The algorithm developed in this study can be used for automatic seizure detection, thereby reducing the burden on clinicians for annotating long-term EEG recordings, while still achieving good performance even with limited labeled data.