BISeizuRe: BERT-Inspired Seizure Data Representation to Improve Epilepsy Monitoring
Abstract
BISeizuRe leverages a BERT-inspired encoder (BENDR) for EEG-based seizure detection. The model follows a two-phase training strategy—pre-training on the large Temple University Hospital EEG corpus to learn general EEG representations and fine-tuning on CHB-MIT. Subject-specific fine-tuning reduces false positives per hour to 0.23 FP/h, 2.5× lower than the baseline, while maintaining competitive sensitivity. The study analyses architecture choices, pre-processing, and post-processing pipelines to deliver robust seizure detection for wearable monitoring.
Key Highlights
- Uses a BERT-inspired encoder (BENDR) to learn generalized EEG representations from large unlabeled data.
- Subject-specific fine-tuning reduces false positives to 0.23 FP/h—2.5× lower than the baseline—while keeping sensitivity high.
- Explores architecture, pre-processing, and post-processing design choices tailored for wearable seizure monitoring.
Thorir Mar Ingolfsson
Postdoctoral Researcher
I develop efficient machine learning systems for biomedical wearables that operate under extreme resource constraints. My work bridges foundation models, neural architecture design, and edge deployment to enable real-time biosignal analysis on microwatt-scale devices.