When the Signal Drops: Sleep Staging from Single-Channel EEG Under Noise and Segment Dropout
Automatic sleep staging has moved out of the sleep laboratory. Single-channel EEG devices are now worn at home, on the forehead, for nights at a time — and the signals they return are not the clean signals of the benchmark literature. Electrodes lift, sweat and motion contaminate epochs, and wireless links drop segments outright. Yet staging models are still validated almost exclusively on clean recordings. We measure what that omission hides. Using 14 first-night Sleep Cassette recordings from Sleep-EDF Expanded (Fpz-Cz, 14,025 scored 30-s epochs, subject-wise train/validation/test split), we train a compact 63k-parameter CNN under two regimes — clean training only, and training with random Gaussian noise and contiguous segment dropout — and evaluate both on held-out subjects under graded corruptions. The cleanly trained model is fragile in a specific way: additive noise at 10 dB SNR cuts its macro-F1 from 0.546 to 0.223, and at 0 dB it scores 0.118, below the accuracy of always predicting the majority class, with N2, N3 and REM recall collapsing to zero. Segment dropout looks benign in accuracy (0.51–0.55 at up to 50% of each epoch zeroed) but quietly destroys N3 (recall 0.00 at 50%). Corruption-augmented training repairs most of this: macro-F1 of 0.614 at 10 dB and 0.526 under 50% dropout, and — on this subject split — it also improves clean-test performance (accuracy 0.658 vs 0.531; Cohen's κ 0.53 vs 0.42), behaving as a regulariser across subjects. Robustness of this kind is nearly free; the cost of skipping it is a model whose failure mode is silence in exactly the conditions home monitoring produces nightly.
Authors
- Saagar Sengupta
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-13
- DOI
- https://doi.org/10.5281/zenodo.24209135
- Primary Topic
- EEG and Brain-Computer Interfaces
- Type
- preprint