Cross-Domain Transfer of EEG to EEG or ECG Learning for CNN Classification Models.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Electroencephalography (EEG) offers high temporal resolution for neurological disorder evaluation but can be inconvenient.
• Deep learning models require extensive data and training time, posing challenges for clinical application.
• Transfer learning presents a potential solution to mitigate these limitations.

Purpose of the Study:

• To investigate the effectiveness of EEG-EEG and EEG-ECG transfer learning strategies for training convolutional neural networks (CNNs).
• To apply these strategies to seizure prediction and sleep staging systems.
• To assess the impact on training time and prediction accuracy.

Main Methods:

• Developed patient-specific CNN models using transfer learning for seizure prediction (interictal and preictal periods).
• Implemented cross-signal transfer learning (EEG-ECG) for sleep staging into five stages.
• Utilized frozen layers in the seizure prediction model to expedite personalization.

Main Results:

• The seizure prediction model achieved 100% accuracy for 7/9 patients with only 40s of personalized training.
• The EEG-ECG transfer learning model for sleep staging surpassed the ECG-only model by ~2.5% in accuracy.
• Training time for sleep staging was reduced by over 50% using transfer learning.

Conclusions:

• Transfer learning effectively reduces training time and enhances accuracy for personalized neurological monitoring.
• This methodology addresses key challenges in deep learning, including data scarcity and variability.
• Transfer learning offers a promising approach for developing efficient and accurate brain-computer interfaces.