Learning-Based Framework for Robust Sleep Stage Labels from Multiple Experts
A new study has come out that presents a method for generating trustworthy sleep stage labels by learning from different expert scorers. This research, which you can find on arXiv under the code 2608.12446, uses the DOD-H and DOD-O datasets that are publicly available. They analyzed EEG (C3-M2) and chin EMG signals, breaking them into 30-second segments and extracting 30 features from each, leading to a total of 60 features. The learning-based hypnogram (LBH) captures how each scorer behaves using confusion matrices from machine-learning models. By normalizing these matrices, they estimate the likelihood of each sleep stage, combining insights from multiple experts to improve the reliability of labels, ultimately benefiting automatic sleep staging assessments.
Key facts
- Study proposes a learning-based framework (LBH) for deriving robust sleep stage labels from multiple experts.
- Uses publicly available DOD-H and DOD-O datasets.
- EEG (C3-M2) and chin EMG signals segmented into 30-s epochs.
- 30 features extracted per modality, 60 total for EEG+EMG.
- Confusion matrices from machine-learning models model scorer behavior.
- Probabilities aggregated across scorers to assign final labels.
- Addresses inter-scorer variability in sleep stage classification.
- Published on arXiv with identifier 2608.12446.
Entities
Institutions
- arXiv