M4 Sleep-EDF Primary-Horizon Pre-Analysis Plan#

This note documents the committed M4 Sleep-EDF execution contract. It is a prospective pre-analysis plan for a future local/public-data run, not a completed Sleep-EDF result.

Scope#

  • Dataset: PhysioNet Sleep-EDF Expanded.

  • Raw data policy: raw PSG and hypnogram files must stay outside the repository.

  • Split/cluster unit: dataset-scoped Sleep-EDF subject ID parsed from the filename.

  • Session boundary unit: subject/night ID parsed from the filename.

  • Primary horizon: horizon 1.

  • Descriptive horizons: horizons 2 and 3.

  • Primary endpoint: RFS bits versus the strongest gated nuisance/trivial baseline.

  • Claim boundary: benchmark-method hardening only. This does not permit clinical, diagnostic, treatment, or model-superiority claims.

Required Gates#

The primary horizon is the only inferential horizon in this protocol. Other horizons are descriptive unless a separate multiplicity-controlled plan is added.

The future Sleep-EDF execution must fail closed if:

  • the local Sleep-EDF root is inside the repository,

  • Sleep-EDF filenames cannot be parsed into subject and night metadata,

  • primary-horizon event patients are below 8,

  • primary-horizon positive events are below 100,

  • the primary RFS confidence interval includes zero,

  • shuffled-target or time-shift controls are too close to primary RFS,

  • the patient-cluster sign-flip permutation is not significant,

  • any nuisance probe exceeds chance plus 0.20.

Provenance Requirements#

The runner writes redacted execution metadata only. It records file names, file hashes, parsed subject IDs, parsed night/session IDs, pair counts, the preregistration hash, the primary-horizon result, and gate failures. It does not record absolute local raw-data paths.

Sources#

  • PhysioNet Sleep-EDF Expanded v1.0.0 documents the public PSG/hypnogram dataset, EDF format, and manual sleep-stage scoring.

  • EEG leakage literature motivates subject/session-aware splits and fail-closed nuisance probes because segment-level random splits can overestimate performance.

  • Human medical intervention evidence, such as randomized-trial meta-analysis of exercise for depression, is relevant only to treatment claims. This M4 feature makes no intervention, depression, clinical, or treatment claim.