DINO-EEG: Montage-Robust Event-Level Seizure Detection in Scalp EEG

This is the authors' abstract. We don't add key points for this paper.

Manual review of long-term clinical electroencephalography (EEG) remains a major bottleneck for large-scale seizure screening in intensive care and epilepsy monitoring settings. Most automated seizure detection methods formulate this task as epoch-level classification, which does not directly provide clinically relevant event boundaries. In addition, many existing systems are sensitive to montage configuration and channel availability, which limits their use in routine clinical practice. We present DINO-EEG, a channel- wise framework for event-level seizure detection in scalp EEG. The method processes each EEG channel independently and directly predicts seizure onset and offset using an object-detection formulation, after which channel- wise detections are aggregated into global seizure events. This design targets two practical goals: direct event-level output and improved robustness to variability in electrode configuration. On the Temple University Hospital EEG Seizure Corpus (TUSZ) v2.0.3, evaluated under the event-based SzCORE (Seizure Community Open-Source Research Evaluation) framework, DINO-EEG achieved an event-based F1 score of 0.641, with a mean sensitivity of 0.663 at 12.5 false positives per 24 hours (FPs/24 h) in seizure subjects, while yielding 4.7 FPs/24 h in seizure-free subjects. The method also remained stable under mild-to-moderate channel dropout and retained seizure detection capability during direct transfer to the CHB-MIT and Siena datasets without retraining. These results suggest that channel- wise event detection is a practical and clinically aligned strategy for automated seizure detection in retrospective EEG review workflows.