Semi-supervised Training Data Selection Improves Seizure Forecasting in Canines with Epilepsy.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Data Science

Background:

• Conventional seizure prediction models use uniform sampling of pre-ictal electroencephalogram (EEG) epochs.
• This assumes a continuous pre-ictal brain state, potentially overlooking complex dynamics.
• Stochastic fluctuations in EEG suggest a need for more sophisticated training data selection.

Purpose of the Study:

• To develop and evaluate a semi-supervised technique for selecting pre-ictal EEG data.
• To improve seizure prediction algorithm performance by identifying more distinguishable pre-ictal epochs.
• To assess the impact of hierarchical clustering on seizure forecasting.

Main Methods:

• Proposed a semi-supervised method utilizing hierarchical clustering.
• Identified optimal pre-ictal data epochs distinguishable from interictal EEG.
• Compared seizure forecasting algorithm performance with and without hierarchical clustering in canine epilepsy models.

Main Results:

• Hierarchical clustering significantly improved overall performance for Time In Warning (TIW) and False Positive Rate (FPR).
• Evaluated on chronic iEEG recordings from six canines with naturally occurring epilepsy.
• Individual dog results for TIW, FPR, and Sensitivity showed variability.

Conclusions:

• Hierarchical clustering is beneficial for overall training data selection in seizure prediction.
• Subject-wise evaluation of the clustering method is recommended for optimal results.
• This clustering approach can optimize forecasting for sensitivity, TIW, or FPR, aiding epilepsy management.