NED-Net is presented as open-source software designed to close an annotation–training loop for rodent EEG event detection. The authors reported a workflow in which candidate events are reviewed by humans and then used as labels to train detectors that run live during ongoing recordings, with periodic retraining as annotations accumulate. The tool aims to address a common problem: automated detectors often perform well on their development datasets but can fail to generalize when laboratory electrodes, animal cohorts, or noise characteristics differ.
The study team framed the approach as a way to evaluate and adapt detectors to local recordings. The preprint states that NED-Net supports continuous integration of human review into model updates; however, the article did not report complete technical details of the model architecture or all training hyperparameters in the abstract and summary provided.
A detector trained on one mouse cohort, using 867 annotated seizures as reported by the authors, produced 171 detections that were then inspected by humans. Data indicated a high precision among those reviewed—93.6%—suggesting the model generated few false positives on those specific detections. Short, high-confidence statements: precision was strong in this limited review set.
When the same detector was applied without modification to a second cohort recorded on the same acquisition system, the authors observed that most reviewed detections were seizures (12 of 13 reviewed). Despite that, per-animal recall varied substantially. The reported median per-animal recall was 10%, and four of 11 animals had no convulsive reference seizure detected by the unchanged detector. Those figures suggest that precision on candidate detections does not guarantee comprehensive event capture across individual animals in a new cohort.
The study team retrained the detector using annotations gathered from the second cohort. Retraining produced a marked change in the detector’s sensitivity to convulsive events: convulsive recall rose from 15% before retraining to 64% after retraining, according to the preprint. The authors also noted that retraining produced at least one reference seizure detected in every animal of the retrained cohort, whereas some animals had been missed entirely by the unchanged model.
These findings indicate that cohort-specific annotations can materially alter recall metrics and animal-level coverage. The source material did not detail the exact number of annotations used for retraining in the second cohort within the abstract, nor did it provide architecture-level descriptions in the summary text, so readers should consult the full manuscript or the linked repository for implementation specifics.
To probe generalization further, the authors withheld a set of recordings from training and evaluated detection. On those withheld recordings, 75% of detections were seizures and the detector identified 14 of 22 reference seizures. By contrast, the original source model and a model trained without inclusion of those animals detected none of those withheld reference seizures. The documented comparison demonstrates that including representative, cohort-specific annotations in training materially improved performance on those held-out recordings.
Short technical caveat: the abstract and summary provided do not enumerate metrics such as false-negative rates per animal, or provide confusion matrices; they report aggregated counts and percentages for reviewed detections and reference seizures. Interested readers should consult the full text and supplementary material for expanded performance tables and evaluation protocols; the authors linked code and data repositories for reproducibility.
The authors have made code and data available via GitHub and Zenodo links supplied in the preprint. Funding sources are listed in the manuscript, and the authors declared no competing interests. The preprint is explicitly identified as not having undergone peer review.
Limitations reported in the source include variable per-animal recall when transferring a detector unchanged to a new cohort and the dependence of detection coverage on cohort-specific retraining. The abstract did not include full model specifications, training hyperparameters, or the detailed annotation workflow in this summary. Those elements appear to be available in the full manuscript and supplementary files linked by the authors; readers should review those materials to reproduce experiments or to adapt NED-Net in their own laboratories.
The article ends with software access information and references to code and data archives. The authors position NED-Net as a practical tool for evaluating and adapting seizure detectors to local rodent EEG recordings, while acknowledging that detector performance depends on the representativeness of annotations used for training and retraining.