Temporal context, not encoder capacity, drives automatic sleep staging

Code, fold definitions and per-fold results for the manuscript:

Temporal context, not encoder capacity, drives automatic sleep staging: a subject-disjoint evaluation of compact models on Sleep-EDF-78.

Milton Vladimir Mamani Calisaya (ORCID: 0000-0002-0676-0989) and Vladimiro Ibañez Quispe (ORCID: 0000-0002-0277-4945), Universidad Nacional del Altiplano, Puno, Peru.

Headline results

Five-fold subject-level cross-validation over all 78 subjects of Sleep-EDF-78 (195,479 thirty-second epochs, each scored exactly once):

Model	Parameters	Accuracy	Cohen's kappa
full	362,085	0.800 +/- 0.023	0.733 +/- 0.029
compact	76,085	0.789 +/- 0.027	0.720 +/- 0.033

What is in this deposit

• folds.json - explicit person-to-fold assignment. Both nights of a subject always share a fold; this file lets you audit the partitions without re-running GroupKFold.
• code/models/ - the per-epoch encoder and the sequence model.
• code/scripts/ - preprocessing (with sleep-period cropping and EDF integrity verification), local training/evaluation, benchmarks, and the scripts that regenerate every figure and table of the manuscript from the JSON results.
• colab/SleepStaging_CV.ipynb - the full GPU experiment battery (cross-validation, encoder/sequence variants, ablation).
• results/ - per-fold metrics, confusion matrices and training histories for every run reported in the manuscript, plus measured efficiency.

Data

The Sleep-EDF Expanded database is not redistributed here. Download it from PhysioNet (https://physionet.org/content/sleep-edfx/1.0.0/); the preprocessing script verifies every EDF file against the record count declared in its own header before use - in our copy, 42 of 160 PSG files were silently truncated downloads until re-fetched.

Reproducing

1. pip install -r code/requirements.txt
2. Download Sleep-EDF Cassette into data/raw/.
3. python code/scripts/preprocess_v2.py
4. python code/scripts/package_for_colab.py
5. Run colab/SleepStaging_CV.ipynb on a GPU (or code/scripts/cv_reduced_local.py on CPU for the compact model).
6. python code/scripts/report_from_colab.py --dir <results> regenerates tables and figures; build_manuscript.py regenerates the manuscript. Every number in the paper is read from these JSON files at build time.

License

Code: MIT. Results and documentation: CC-BY-4.0.