Lightweight CNN-based seizure classification via leveraging chimera states in iEEG recordings
Epileptic seizures, a chronic brain disorder characterized by groups of neurons sending incorrect signals, result in recurring seizures associated with chimera states—the simultaneous presence of synchronized and desynchronized neural activity. These states provide a potential tool for predicting an...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-09-01
|
| Series: | Results in Engineering |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S2590123025020729 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Epileptic seizures, a chronic brain disorder characterized by groups of neurons sending incorrect signals, result in recurring seizures associated with chimera states—the simultaneous presence of synchronized and desynchronized neural activity. These states provide a potential tool for predicting and detecting epileptic seizures. Effective detection and classification of chimera states in intracranial electroencephalography (iEEG) signals are critical for improving patient outcomes and deepening our understanding of neurological disorders. A major challenge is developing a universally applicable method that accommodates diverse electrode setups across patients. In this study, we propose the Chimera Synchronization Matrix-based Network (CSMNet), a novel algorithm that transforms iEEG signals into 2D images, capturing spatial-temporal dynamics with reduced computational complexity. These images are processed by a streamlined convolutional neural network (CNN) framework, which classifies iEEG recordings into pre-ictal, ictal, and post-ictal events with robust patient-independent performance. Trained on only 10 epochs with a limited dataset from the SWEC-ETHZ database, our CNN achieved an accuracy of 96.67% on short-term iEEG recordings (excluding one patient) and 95.2% on long-term recordings. Notably, the false detection rate (FDR) was 0% for 5 out of 14 patients in the short-term dataset and 4 out of 18 patients in the long-term dataset. Compared to prior studies, our approach uses fewer parameters (17,083) and epochs, enhancing computational efficiency. With a sensitivity of 92.12% (short-term) and 87.28% (long-term), and specificity of 95.99% (short-term) and 93.33% (long-term), this framework offers significant promise for clinical diagnostics, real-time monitoring, and personalized epilepsy treatment planning. |
|---|---|
| ISSN: | 2590-1230 |