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

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Neurology

Background:

• Epilepsy seizure forecasting remains challenging, especially for long-term predictions.
• Existing models often rely on raw electroencephalography (EEG) data, limiting their effectiveness for long-horizon forecasting.
• Responsive neurostimulation (RNS) systems generate valuable biomarkers that can potentially improve seizure prediction.

Purpose of the Study:

• To develop and evaluate SeizureFormer, a novel Transformer-based model for long-horizon seizure risk forecasting (1-14 days).
• To leverage structured biomarkers from RNS systems, specifically interictal epileptiform activity (IEA) and long episodes (LE), for improved prediction accuracy.
• To compare SeizureFormer's performance against existing statistical, classical machine learning, and deep learning models.

Main Methods:

• Developed SeizureFormer, a Transformer-based deep learning model integrating multi-scale CNN patch embedding, cross-variable temporal convolution, and squeeze-and-excitation attention.
• Utilized structured biomarkers (IEA and LE) extracted from RNS systems for model training and testing.
• Evaluated the model across five patients with multiple prediction windows ranging from 1 to 14 days.

Main Results:

• SeizureFormer achieved state-of-the-art performance, with a mean ROC AUC of 79.44% and mean PR AUC of 76.29% across patients and prediction windows.
• The model demonstrated superior generalizability compared to baseline models, particularly under conditions of class imbalance.
• Successfully predicted seizure-related events 1 to 14 days in advance.

Conclusions:

• SeizureFormer offers a significant advancement in long-horizon seizure risk forecasting using RNS biomarkers.
• The model's ability to capture both short-term fluctuations and long-term seizure cycles enhances prediction accuracy.
• This technology enables actionable multi-day forecasting, paving the way for personalized and proactive interventions in epilepsy management.