A novel consistency-based training strategy for seizure prediction.

Area of Science:

• Neurology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Epilepsy seizure prediction is crucial for patient well-being.
• Deep learning models show promise but struggle with generalization due to EEG's non-stationary nature and data scarcity.
• Existing methods primarily focus on model architecture, neglecting stability against perturbations.

Purpose of the Study:

• To propose a novel consistency-based training strategy to enhance the generalization ability of deep learning models for seizure prediction.
• To improve the robustness of seizure prediction models against small input perturbations.

Main Methods:

• A consistency-based training strategy is introduced, enforcing consistent outputs for perturbed inputs.
• Stochastic augmentations are applied during training to create input variations.
• A consistency constraint penalizes differences between current and previous outputs to enhance generalization.

Main Results:

• The proposed strategy was implemented in two state-of-the-art models (STFT CNN, Multi-view CNN) on scalp and intracranial EEG datasets.
• Significant performance improvements were observed for both implemented models.
• The strategy increased sensitivity by 7.1% and reduced false prediction rate by 0.12/h on one baseline, and improved AUC by 0.020 on another.

Conclusions:

• The proposed consistency-based training strategy is effective in enhancing seizure prediction model performance.
• The strategy is easy to implement and offers a new solution for improving seizure prediction accuracy.
• This approach addresses the critical need for robust and generalizable deep learning models in epilepsy management.