Sleep-stage classification
Updated
Project: completed
I completed an internal study of how sleep-stage classifiers generalize to people they were not trained on. The implementation and reserved-person evaluation are complete. Publication and deployment remain pending; this page is a concise project record, not a paper or an inference service.
Research question
An EEG recording changes over time, and recordings from the same person share characteristics. I studied whether neighboring segments improve sleep-stage prediction beyond a matched model that sees only the current segment, while keeping people separated during development and evaluation.
Data and method
The study uses the Sleep Cassette subset of Sleep-EDF Expanded. Recordings from the same person stay in the same partition. The task assigns a sleep-stage label to each complete 30-second segment, separating wake, N1, N2, combined deep sleep and REM. Preprocessing, stage mapping and evaluation choices are recorded in the research repository.
The implemented comparisons include spectral-feature baselines, a compact convolutional network, matched current-segment and temporal-context models, and temporal decoding controls. The internal final evaluation uses a reserved group of people after development decisions were frozen.
Evaluation and release status
Internal evaluation is complete. Numerical results, model artifacts and detailed figures are being held for a later release. There is no public paper, model download, repository link or deployed inference API on this page.
The primary evaluation gives each person equal weight when averaging five-class macro-F1, rather than allowing longer recordings to dominate. Evaluation also considers performance by stage and errors near stage transitions. An overall score can conceal adverse effects for particular stages or temporal boundaries; those limitations belong alongside any eventual result.
Limits
This is evaluation within one source cohort. It does not establish generalization to another hospital, recording system or population, clinical usefulness, or peer-reviewed findings. Some temporal comparisons use future context and therefore describe offline analysis, not a real-time system.
Next decision
Preserve the completed evaluation and its artifacts. Publication, external-cohort validation and deployment are separate future decisions. The current study is complete; those steps remain pending.
Project workflow
Figures and evaluation
Evaluation exports are not available on this site.
Project links
No public repository or demo is linked here.