Authors :
Presenting Author: Gagan Acharya, PhD Candidate – University of California, Riverside
Erin Conrad, MD, MA – University of Pennsylvania
Kathryn Davis, MD – Center for Neuroengineering and Therapeutics and Penn Epilepsy Center, Department of Neurology, University of Pennsylvania
Erfan Nozari, PhD – University of California, Riverside
Rationale:
Seizures evolve through complex, multiscale brain dynamics, yet most existing seizure forecasting methods treat iEEG features as temporally independent samples, potentially limiting predictive accuracy. To address this, we investigate whether explicitly modeling the dynamics of both iEEG features and seizure risk can enhance forecasting performance. Furthermore, to simulate constraints imposed by implanted neurostimulation systems such as the RNS system, which operate using a limited number of channels (typically 4–8) and rely on a small set of simple signal features (line length, variance, half-wave counts, band power), we emulate a virtual RNS environment using continuous iEEG recordings. This enables systematic, pseudo-prospective (PP) evaluation of forecasting algorithms under clinically realistic conditions while preserving temporal continuity.
Methods:
We analyzed continuous multi-day (7 ± 2 days) intracranial EEG (iEEG) EMU recordings from 30 subjects with drug-resistant mesial temporal lobe epilepsy (MTLE), publicly available through the IEEG.org portal. To approximate the constraints of an implanted RNS system, we emulated a virtual RNS-like environment by restricting analysis to four channels located near the mesial temporal lobe (MTL), and by extracting a limited set of low-compute signal features from short (~10-second) non-overlapping time windows. We trained baseline seizure classifiers using traditional i.i.d. feature inputs and augmented them with autoregressive (AR) models that capture dynamics of (1) classification features, (2) seizure likelihood estimates, and (3) a hybrid of both. Performance was evaluated pseudo-prospectively via PP-AUC and time-in-warning (TiW) metrics, cross-validated per subject.
Results:
Dynamical augmentation significantly improved PP-AUC over the baseline predictor (Wilcoxon signed-rank test, p < 0.05), with the hybrid model performing best. Improvements were robust across prediction horizons, peaking at ~15 minutes before seizure onset. Notably, incorporating feature and risk dynamics reduced the time-in-warning (TiW) needed to detect all seizures at 100% sensitivity, indicating more efficient and timely seizure forecasting. Interictal duration preceding seizures was negatively correlated with TiW (Spearman’s ρ < 0, p < 0.05), indicating easier prediction after longer seizure-free periods. Electrode location analysis revealed that while mesial temporal contacts (near SOZ) were generally optimal for prediction, a third of subjects showed superior forecasting using lateral temporal channels, highlighting the importance of individualized electrode selection.