Addressing data limitations in seizure prediction through transfer learning.

Area of Science:

• Neurology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Patient-specific seizure prediction models are crucial but limited by rare seizure events.
• Optimizing these models requires substantial patient data, which is often scarce.

Purpose of the Study:

• To investigate the efficacy of transfer learning using external patient data to enhance patient-specific seizure prediction models.
• To address the data scarcity challenge in developing robust seizure prediction systems.

Main Methods:

• A deep convolutional autoencoder was trained on electroencephalogram (EEG) data from 41 patients (EPILEPSIAE database).
• Transfer learning was applied by adding bidirectional long short-term memory and classifier layers, optimized for 24 individual patients (Universitätsklinikum Freiburg).
• The pre-trained encoder acted as a fixed feature extractor for patient-specific model optimization.

Main Results:

• Seizure prediction models optimized with pre-trained weights showed approximately four times fewer false alarms.
• The models maintained seizure prediction accuracy and achieved 13% more validated patients.
• Transfer learning resulted in more stable and faster training, saving computational resources.

Conclusions:

• Transfer learning offers a significant advancement for seizure prediction, overcoming data limitations.
• This method provides more efficient, stable training and conserves computational resources.
• Transfer learning facilitates easier data sharing due to fewer ethical and storage constraints.