Generalizable epileptic seizures prediction based on deep transfer learning.

Area of Science:

• Neurology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Epileptic seizures pose significant challenges for patient well-being and require effective prediction methods.
• Early seizure prediction enables timely intervention, potentially preventing adverse events through medication.

Purpose of the Study:

• To investigate the efficacy of deep transfer learning models for predicting epileptic seizures.
• To identify optimal models and parameters for both patient-dependent and patient-independent seizure prediction.

Main Methods:

• Extracted 22 features from 5-second electroencephalogram (EEG) segments.
• Developed tensor inputs for deep transfer learning models, including ImageNet convolutional networks (Xception, MobileNet-V2) and classifiers (Fully Connected).
• Evaluated models using varying pre-ictal state durations (10, 20, 30, 40 minutes) and patient-dependent/independent testing.

Main Results:

• The Xception model with a Fully Connected classifier achieved 98.47% sensitivity and a 0.031 h⁻¹ False Prediction Rate (FPR) for patient-dependent prediction over a 40-min pre-ictal state.
• The MobileNet-V2 model with a Fully Connected classifier demonstrated patient-independent prediction with 98.39% sensitivity and 0.029 h⁻¹ FPR for a 40-min pre-ictal scheme.

Conclusions:

• Deep transfer learning models show high accuracy in predicting epileptic seizures.
• Patient-independent seizure prediction is feasible and highly accurate, offering significant potential for clinical application.