Continual learning for seizure prediction via memory projection strategy.

Area of Science:

• Machine Learning
• Neuroscience
• Biomedical Engineering

Background:

• Epilepsy prediction models often fail in real-world dynamic data scenarios.
• Catastrophic forgetting (CF) is a major challenge when learning from continuously changing electroencephalogram (EEG) data.
• Existing machine learning algorithms are typically designed for static, offline datasets.

Purpose of the Study:

• To implement and evaluate a continual learning (CL) strategy, Memory Projection (MP), for dynamic epilepsy prediction.
• To address the problem of catastrophic forgetting (CF) in machine learning models trained on evolving EEG data.
• To develop a robust epilepsy prediction model suitable for real-time clinical application.

Main Methods:

• Implemented a continual learning (CL) strategy called Memory Projection (MP).
• Integrated Regularization Loss Reconstruction and Matrix Dimensionality Reduction Algorithms into the MP core.
• Trained and evaluated the MP strategy on sequential and multi-center EEG datasets.

Main Results:

• MP demonstrated excellent performance in sequential seizure prediction with forgetting rates below 5% for accuracy and sensitivity.
• When applied to multi-center datasets, MP achieved very low forgetting rates (0.65% for accuracy, 1.86% for sensitivity).
• Ablation studies confirmed MP's minimal storage and computational requirements, highlighting its practical clinical potential.

Conclusions:

• Memory Projection (MP) is an effective continual learning strategy for epilepsy prediction using EEG data.
• MP significantly mitigates catastrophic forgetting, enabling robust performance in dynamic and multi-center clinical settings.
• The low computational and storage costs of MP make it a practical solution for real-time seizure prediction.