Hybrid Network for Patient-Specific Seizure Prediction from EEG Data.

Area of Science:

• Neurology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Epilepsy affects millions globally, with drug-resistant cases significantly impacting quality of life.
• Accurate seizure prediction is crucial for managing epilepsy, especially in drug-resistant forms.
• Current deep learning models, particularly single convolution networks, struggle with capturing long-term dependencies in EEG data.

Purpose of the Study:

• To develop an advanced deep learning model for accurate seizure prediction.
• To address the limitations of single convolution models in capturing global and long-term EEG features.
• To improve the performance of seizure prediction systems for patients with drug-resistant epilepsy.

Main Methods:

• A hybrid deep learning model, STCNN, integrating Swin transformer (ST) and 2D convolutional neural network (2DCNN) was developed.
• Time-frequency features were extracted using short-term Fourier transform (STFT) as input for the STCNN model.
• ST blocks captured global information and long-term dependencies, while 2DCNN blocks focused on local and short-term features.

Main Results:

• The STCNN model achieved an average seizure prediction sensitivity of 92.94%.
• The area under the ROC curve (AUC) reached 95.56%, indicating high prediction accuracy.
• A low false positive rate (FPR) of 0.073 was recorded, demonstrating the model's reliability.

Conclusions:

• The proposed STCNN model effectively combines global and local feature extraction for superior seizure prediction.
• This hybrid approach significantly improves upon traditional single convolution models for EEG-based seizure forecasting.
• The STCNN model shows strong potential for clinical application in improving the lives of epilepsy patients.