Reducing False Alarms in Wearable Seizure Detection With EEGformer: A Compact Transformer Model for MCUs
EEGformer hardware-aware deployment
Abstract
Long-term seizure monitoring in wearables suffers from high false-alarm rates. EEGformer is a compact transformer-based detector tailored for low-power microcontroller units that operates directly on raw temporal EEG channels. Hardware-aware optimization enables EEGformer to detect 73% of seizures on the CHB-MIT dataset with only 0.15 false positives per hour while reducing detection latency by 20%. Deployment on the GAP9 MCU performs inference in 13.7 ms at 0.31 mJ per inference, demonstrating practical suitability for wearable seizure detection devices with multi-day autonomy.
Key Highlights
- Tiny transformer architecture tailored for raw temporal EEG and low-channel wearable acquisition.
- Detects 73% of seizures with 0.15 false positives per hour and reduces detection latency by 20% on CHB-MIT.
- Hardware-aware implementation achieves 13.7 ms inference latency at 0.31 mJ on GAP9, enabling multi-day wearable deployment.
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.