{"id":"arxiv:2605.10688","kind":"preprint","title":"DANCE: Detect and Classify Events in EEG","authors":["Jarod Lévy","Hubert Banville","Jérémy Rapin","Jean-Remi King","Thomas Moreau","Stéphane d'Ascoli"],"author_count":6,"year":2026,"published":"2026-05-11","venue":null,"cited_by":null,"doi":null,"arxiv_id":"2605.10688","nct_id":null,"source_url":"https://arxiv.org/abs/2605.10688","archive_url":"https://space-kitty.com/papers/arxiv:2605.10688","updated":"2026-05-11","pmid":null,"pmcid":null,"openalex_id":null,"pdf_url":"https://arxiv.org/pdf/2605.10688","abstract":"Event identification in continuous neural recordings is a critical task in neuroscience. Decoding in EEG is dominated by classifying windows aligned to known event onsets. However, while available in controlled experiments, such onsets are absent in continuous real-world monitoring. Here, we introduce DANCE, a deep learning pipeline that frames neural decoding as a set-prediction problem and jointly detects and classifies events directly from raw, unaligned signals. Evaluated separately on ten datasets curated from the literature with a wide variety of event types (ranging from milliseconds to minutes in duration), our model outperforms existing methods on a broad range of cognitive, clinical and BCI tasks. This single architecture establishes a new state of the art in the competitive task of seizure monitoring and matches the accuracy of onset-informed models for BCI tasks. Overall, our method marks a step towards end-to-end asynchronous neural decoding models","abstract_license":"CC0-1.0 (arXiv descriptive metadata)","license":null,"is_open_access":false,"retracted":false,"peer_reviewed":false,"categories":["cs.LG","eess.SP"],"details":null,"provenance":"arxiv","attribution":"Abstract: arXiv descriptive metadata (CC0). Thank you to arXiv for use of its open access interoperability. This service was not reviewed or approved by, nor does it necessarily express or reflect the policies or opinions of, arXiv.","topics":[],"cited_in_dossiers":[]}