Epilepsy seizure prediction based on ViT-2DCNN spatio-temporal fusion model

Area of Science:

• Neurology
• Biomedical Engineering
• Machine Learning

Background:

• Epilepsy is a chronic neurological disorder marked by recurrent seizures.
• Accurate seizure prediction is crucial for timely medical intervention.
• Current models struggle with patient-specific electroencephalogram (EEG) data, limiting prediction reliability.

Purpose of the Study:

• To develop an improved seizure prediction model.
• To enhance prediction accuracy and robustness by integrating time-frequency and spatial EEG information.
• To overcome limitations of single-modality prediction approaches.

Main Methods:

• Proposed a spatio-temporal fusion model (ViT-2DCNN) for seizure prediction.
• Introduced an entropy distribution map as a spatial-modality input, preserving electrode topology and reflecting brain complexity.
• Combined a Vision Transformer (ViT) for global time-frequency analysis and a 2DCNN with spatial attention for local patterns, using a gated fusion module.

Main Results:

• Achieved high performance on the CHB-MIT dataset: Accuracy 97.95%, Sensitivity 98.36%, Specificity 97.55%, F1-score 97.98%.
• Six subjects achieved 100% accuracy.
• Demonstrated reliable lower-bound efficacy with the lowest accuracy exceeding 92.37% across all subjects.

Conclusions:

• Fusing time-frequency and spatial entropy features captures richer spatio-temporal EEG characteristics.
• The ViT-2DCNN model overcomes limitations of single-modality methods.
• The findings indicate strong potential for clinical application in epilepsy seizure prediction.