ALARM-Net: False-Alarm Suppression for EEG Seizure Detection

ALARM-Net is an event-level false-alarm suppression framework for automated seizure detection on the Temple University Hospital (TUH) EEG Corpus. Automated seizure detectors are often unusable in practice because they raise too many false alarms per 24 hours. ALARM-Net adds a second-stage event scoring model on top of a window-level detector and reduces false alarms per 24 h by 58.9% on the development split and 76.3% on the held-out evaluation split, at a cost of 7.5 to 7.8 percentage points of strict event sensitivity. Manuscript submitted to MDPI Applied Sciences (2026), co-authored with Prof. Abdulhamit Subasi at the University at Albany. Stack: Python, PyTorch, MNE, SLURM on a DGX A100 cluster.