SeizureFormer: A Multi-Scale Transformer for Seizure Risk Forecasting from RNS-Derived Biomarkers.

Abstract

We present SeizureFormer, a Transformer-based model for long-horizon seizure risk forecasting (1-14 days) using structured biomarkers-interictal epileptiform activity (IEA) and long episodes (LE)-extracted from responsive neurostimulation (RNS) systems. Unlike prior models based on raw scalp EEG, SeizureFormer leverages stable RNS biomarkers and integrates multi-scale CNN patch embedding, cross-variable temporal convolution, and squeeze-and-excitation attention to capture both short-term fluctuations and long-term seizure cycles. Tested across five patients and multiple prediction windows (1-14 days), SeizureFormer achieved state-of-the-art performance with mean ROC AUC of 79.44% and mean PR AUC of 76.29% across five patients and four prediction windows. Compared to statistical, classical ML, and deep learning baselines, it demonstrates superior generalizability under class imbalance. Clinically, it enables actionable multi-day forecasting, supporting personalized and proactive intervention in epilepsy care by forecasting seizurerelated events 1 to 14 days ahead.