Epilepsy seizure prediction with few-shot learning method.

Area of Science:

• Neurology
• Artificial Intelligence
• Machine Learning

Background:

• Epileptic seizure prediction is crucial for patient management.
• Accurate prediction is challenging for patients with infrequent seizures due to difficulties in recording preictal data.
• Existing generalized methods lack accuracy for this patient group.

Purpose of the Study:

• To develop a few-shot learning approach for epileptic seizure prediction.
• To improve the accuracy of preictal period diagnosis, especially for patients with late-occurring seizures.
• To reduce the data collection burden by leveraging prior knowledge.

Main Methods:

• A convolutional neural network (CNN) model was employed.
• A few-shot learning strategy was implemented, utilizing prior knowledge and a small number of new patient samples.
• The model was evaluated on three patients from the CHB-MIT database.

Main Results:

• Achieved high sensitivity (95.70% for 10-min SPH/20-min SOP, 98.52% for 5-min SPH/25-min SOP).
• Demonstrated a low false prediction rate (0.057/h and 0.045/h, respectively).
• The few-shot learning method outperformed generalized methods in accuracy.

Conclusions:

• The proposed few-shot learning CNN effectively predicts epileptic seizures, particularly for challenging patient cases.
• This approach enhances prediction accuracy and efficiency by minimizing data requirements.
• The method offers a promising advancement in personalized seizure prediction technology.