NED-Net: an annotation-training loop for rodent EEG event detection

Abstract

Long-term EEG monitoring is central to rodent epilepsy research, and manual review is laborious because recordings span hundreds to tens of thousands of animal-hours. Published automated detectors perform well on their development datasets but are typically fixed models that may not adapt to another laboratory's electrodes, animals, or noise conditions. We introduce NED-Net, open-source software for training and maintaining seizure detectors. It uses reviewed candidate events as training labels, runs detection live during recording, and retrains as annotations accumulate. A detector trained on one mouse cohort (867 annotated seizures) achieved 93.6% precision in human review of 171 detections. Applied unchanged to a second cohort on the same acquisition system, its detections were mostly seizures (12 of 13 reviewed), but per-animal recall varied substantially (median 10%), with no convulsive reference seizure detected in four of 11 animals. Retraining with annotations from the second cohort raised convulsive recall from 15% to 64%, with at least one reference seizure detected in every animal. On recordings withheld from training, 75% of detections were seizures and 14 of 22 reference seizures were detected, whereas the source model and a model trained without these animals detected none. NED-Net supports evaluating and adapting detectors to local recordings.