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

Objective. Most existing seizure prediction approaches rely on cohort-based models or assume a single model suffices per patient, overlooking clinical and electrophysiological variability across seizures. This study aims to overcome these limitations by introducing a subject-specific seizure prediction framework that models intra-subject heterogeneity by identifying and clustering seizure-specific preictal patterns using long-term intracranial EEG (iEEG) recordings collected over a one to two week duration. Approach: Absolute, relative, and ratio power spectral density features are extracted from twelve frequency bands, and the minimum uncertainty and sample elimination algorithm is used for unsupervised feature selection on a per-seizure basis. Weighted aggregation is then applied to form seizure-specific feature sets. Seizures are grouped into clusters based on feature similarity, and separate classifiers are trained for each cluster. Model predictions are combined using a grid optimized k-of-N voting strategy. Evaluation is conducted on long-term iEEG recordings from ten patients using cross-validation across seizure-containing sessions. Main results. When clustering is applied, the mean sensitivity across subjects is improved from 89.17% to 98.54%, while the mean FPR is reduced from 1.15/day to 0.62/day. Additionally, the median number of features required per subject decreased from 22 to 14, reflecting a 36.4% reduction in model complexity. Finally, in 72.5% of subject-folds, the number of algorithm-identified clusters equaled or exceeded the clinically annotated seizure types, with a linear trend indicating latent electrophysiological variability beyond clinical labels. Significance. These findings highlight the value of modeling seizure diversity within individuals and support the development of more personalized and interpretable seizure forecasting systems.