DCDA-Net: A depthwise convolution dilated attention network with temporal weighted loss for seizure prediction

The sudden and unforeseeable nature of seizures poses serious threats to patients' safety and quality of daily life, making the development of highly accurate and reliable seizure prediction methods of significant clinical importance. To this end, this study proposed a seizure prediction model by integrating depthwise convolution and a novel dilated attention mechanism, named DCDA-Net. Depthwise convolution was first employed to extract features from each EEG channel independently, preserving spatial channel information for subsequent seizure evolution analysis. Innovatively combining dilated convolution with self-attention mechanisms, a dilated attention module was devised to efficiently capture long-range temporal dependencies. Furthermore, a temporal weighted loss function was introduced to significantly enhance the model's ability to discriminate critical preictal temporal features. Validation results on the CHB-MIT and SH-SDU clinical databases demonstrated that the model achieved a segment-based accuracy of 97.56% and an event-based sensitivity of 97.69% with a false prediction rate of 0.0232/hour on the CHB-MIT database, while attaining a segment-based accuracy of 96.93% and an event-based sensitivity of 96.05% on the SH-SDU database. Additionally, channel importance analysis confirmed significant spatial consistency with the actual epileptogenic zones. This study offered an efficient solution for patient-specific seizure prediction, with substantial potential for further clinical assessment.