Temporal context and transition decoding for sleep staging in unseen participants on Sleep-EDF
We evaluated established temporal methods for offline sleep staging of reference-eligible epochs in Sleep-EDF Expanded. Development used 62 participants and final evaluation used 16 unseen participants, with each participant's nights kept together. Linear classifiers shared a frozen encoder and fitting settings but received either the target representation alone or the target together with its preceding and following neighbors. The primary endpoint was participant-average five-class macro-F1 under participant-and-stage training weights, pooling nights and averaging three seed-specific scores within each participant.Local context increased macro-F1 from 0.7119 to 0.7372: a paired gain of 0.0253 (conditional 95% participant-bootstrap interval, 0.0212–0.0292), with all 16 participants improving. Under this weighting policy, the full-transition decoder scored higher overall (0.7479), but lower on reference-defined interior-transition epochs. Against the center-only classifier, context improved stable-interior performance; the mean transition effect was near zero, with uncertainty spanning both signs. Pooled stage metrics showed higher N1 and REM F1 but lower N3 recall and F1.In this fixed model panel, outperforming a decoder at transitions did not imply an improvement over center-only inputs at transitions. The comparison provides participant-level evidence within one data source; it does not introduce a new temporal method. Active capacity differed between the heads, and external-cohort and clinical performance remain untested.This record contains the preprint only. Study code, saved predictions and trained checkpoints are not included. Source Sleep-EDF Expanded recordings, annotations and subject metadata are available separately from PhysioNet under their original terms. The paper and original figures are licensed under CC BY 4.0. This work has not undergone peer review.
Authors
- Satish Bajagain
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-14
- DOI
- https://doi.org/10.5281/zenodo.24212811
- Primary Topic
- EEG and Brain-Computer Interfaces
- Type
- preprint