Patient-specific long-term seizure prediction via multi-model classification.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Data Science

Background:

• Existing seizure prediction models often fail to account for individual patient variability.
• Cohort-based or single-model approaches overlook the heterogeneity of seizures within a single patient.

Purpose of the Study:

• To develop a subject-specific seizure prediction framework addressing intra-subject heterogeneity.
• To model seizure diversity by clustering seizure-specific preictal patterns from long-term intracranial EEG (iEEG) data.

Main Methods:

• Extracted power spectral density features from twelve frequency bands.
• Employed unsupervised feature selection and weighted aggregation for seizure-specific feature sets.
• Utilized clustering to group seizures and trained separate classifiers per cluster, combined with a k-of-N voting strategy.

Main Results:

• Mean sensitivity improved from 89.17% to 98.54%, and mean false positive rate (FPR) decreased from 1.15/day to 0.62/day.
• Model complexity reduced by 36.4% (median features from 22 to 14).
• Identified seizure clusters often exceeded clinically annotated seizure types, revealing latent electrophysiological variability.

Conclusions:

• Modeling seizure diversity within individuals is crucial for advancing seizure forecasting.
• This approach supports the development of more personalized and interpretable epilepsy management systems.