A Multi-Scale Spatiotemporal Capsule Network for Epilepsy Seizure Detection

The electroencephalogram (EEG) signal is pivotal for the expert diagnosis of epilepsy in patients. However, experiential bias among these experts can sometimes lead to inconsistent judgment outcomes. To mitigate this, it is imperative to establish a comprehensive algorithm to aid in resolving this issue. This study introduces a novel deep learning architecture that integrates convolutional neural network (CNN) and capsule neural network (CapsNet) to provide end-to-end epilepsy diagnosis. Using CNN facilitates the extraction of both temporal and spatial information from EEG data. Subsequently, CapsNet analyses the resultant hybrid feature vector, enabling the categorization of interictal and ictal cases. The model was validated through its application to the CHB-MIT and SIENA public datasets for validation purposes. In the domain of seizure event detection, our model demonstrates the accuracy of 99.12% and 99.54%, the sensitivity of 89.97% and 90.93%, specificity of 99.27% and 99.59%, and F1 score of 77.15% and 77.27%, respectively. Furthermore, The CHB-MIT dataset was utilized for seizure onset detection, resulting in a sensitivity rate of 99.13% and a time delay of 9.28 s. Collectively, these results underline that our model adeptly extracts meaningful features from EEG datasets, ensuring accurate judgments.