Dynamic spatiotemporal graph neural network for patient-independent seizure detection with scalp EEG

Abstract The application of deep learning techniques for detecting epileptic seizures from electroencephalogram (EEG) signals has demonstrated significant potential in clinical practice. However, current seizure detection algorithms typically focus on features extracted from individual EEG channels, neglecting the spatial relationships among multiple channels. Existing algorithms are also limited in their ability to adapt to variations in the number of EEG channels and exhibit poor performance in patient-independent seizure detection. To address these challenges, this paper introduces a novel approach that constructs a brain effective connectivity graph from EEG time series using an adaptive directed transfer function. Subsequently, we propose a dynamic spatiotemporal graph neural network (DSTGNN) for detection of epileptic seizure. This model effectively integrates spatial and temporal information to accurately model the temporal progression of epileptic seizure propagation, thereby capturing both spatial and temporal dependencies. Patient-independent experiments conducted on the CHB-MIT and TUSZ datasets achieved average accuracy, recall, precision, and F1-score of 99.23%, 98.37%, 98.04%, and 98.01%, respectively. Experiment results demonstrate that the DSTGNN model not only adapts effectively to variation in the number of EEG channels but also maintains high accuracy in patient-independent seizure detection, outperforming other state of the art models. To the best of our knowledge, this is the first successful application of DSTGNN for seizure detection.