Real-Data Case Study: Sleep-EDF N3 vs Wake Honest Audit¶

Abstract¶

We apply SPO's detector-agnostic auditor to a simple slow-wave detector on a public polysomnogram from PhysioNet Sleep-EDF Expanded. The detector scores each 30-second epoch by the normalized delta-band Hilbert envelope and is asked to separate expert-scored N3 (slow-wave sleep) epochs from Wake epochs at a matched false-alarm rate. On subject SC4001, first night, the detector fires on 95.9 % of N3 epochs while holding the Wake false-alarm rate at 10.0 %, with a permutation p-value < 0.001. The result is sealed into a content-addressed audit record and guarded by an integrity test that needs only the committed artefacts, not the raw recording.

The question¶

Sleep staging is a classic cyclic-state classification problem: the EEG oscillator field moves through Wake → N1 → N2 → N3 → REM and back. A common heuristic in SPO's exploratory sleep-staging script is that cross-band Kuramoto-R phase coherence can distinguish stages. We ask a stricter, operational question: at a fixed false-alarm budget calibrated on Wake epochs, does a detector score N3 epochs higher than the matched false-alarm rate explains? This is the same honest-evaluation moat used in the early-warning study, applied to a sleep domain where the "transition" is the onset of slow-wave sleep.

Data¶

The recording is subject SC4001, first night, from the PhysioNet Sleep-EDF Expanded corpus:

SC4001E0-PSG.edf — polysomnogram at 100 Hz, including EEG Fpz-Cz.
SC4001EC-Hypnogram.edf — expert sleep-stage annotations at 30-second resolution.

Raw files are citation-only and are not redistributed. Cite:

Kemp B, Zwinderman AH, Tuk B, Kamphuisen HAC, Oberye JJL. Analysis of a sleep-dependent neuronal feedback loop: the slow-wave microcontinuity of the EEG. IEEE Transactions on Biomedical Engineering 47(9):1185–1194, 2000.
Goldberger AL, Amaral LAN, Glass L, Hausdorff JM, Ivanov PCh, Mark RG, Mietus JE, Moody GB, Peng C-K, Stanley HE. PhysioBank, PhysioToolkit, and PhysioNet: components of a new research resource for complex physiologic signals. Circulation 101(23):e215–e220, 2000.

Methods¶

Detector¶

The detector extracts the Fpz-Cz channel, applies a 0.5–4 Hz Butterworth band-pass, computes the Hilbert envelope of both the delta-filtered signal and the raw broadband signal, and returns, for each 30-second epoch:

score = mean(|H(delta_filter(eeg))|) / mean(|H(eeg)|)

This is the fraction of the EEG's instantaneous broadband power concentrated in the delta band. N3 is defined clinically by high-amplitude delta slow waves, so this is an amplitude-envelope detector rather than a phase-coherence detector.

Why not Kuramoto R?¶

During development we evaluated the cross-band Kuramoto-R score used by the exploratory experiments/sleep_staging_eeg.py script. On this recording it did not separate N3 from Wake at the matched-false-alarm bar (p ≈ 0.7). The honest result therefore uses the oscillator observable that actually carries the signal: the slow-wave amplitude envelope.

Audit protocol¶

Events: 220 expert-scored N3 epochs.
Nulls: 1997 expert-scored Wake epochs from the same recording.
Calibration: choose the smallest score threshold that holds the Wake false-alarm rate at or below 10 %.
Significance test: 10 000 label permutations (fixed seed 42) of the pooled N3 + Wake alarm outcomes; one-sided p-value for the observed N3 alarm count.

The entire audit is performed by scpn_phase_orchestrator.evaluation.audit_detector and sealed with seal_detector_audit, producing a SHA-256 content hash over the corpus provenance and verdict.

Results¶

Quantity	Value
Epochs analysed	2650
N3 epochs (events)	220
Wake epochs (null)	1997
Mean N3 score	0.837
Mean Wake score	0.654
Matched threshold	0.753415
Target false alarm	0.100
Achieved false alarm	0.100
N3 detection rate	0.959
Permutation p-value	< 0.001
Beats chance (α = 0.05)	yes

The normalized delta envelope cleanly separates the two classes at the matched-false-alarm operating point.

Reproduction¶

Install SPO with the EDF ingestion extra:

pip install "scpn-phase-orchestrator[eeg]"

Obtain the two EDF files from PhysioNet and run:

python bench/sleep_staging_sleepedf.py \
  SC4001E0-PSG.edf \
  SC4001EC-Hypnogram.edf \
  examples/real_data/sleepedf_staging

The script is deterministic, so regenerating the output reproduces the committed sleepedf_n3_vs_wake_audit.json and sleepedf_n3_vs_wake_summary.json.

Provenance and integrity¶

The sealed audit record lives at examples/real_data/sleepedf_staging/sleepedf_n3_vs_wake_audit.json. Its content_hash is:

836b9deda96455b31734c319cb3f30e87fb1ec005fe4a397a555619b57e690d0

tests/test_sleepedf_staging_evidence.py recomputes the content seal from the recorded fields and pins the SHA-256 digests of the citation-only source EDFs, so the result can be guarded without redistributing the raw data.

Scope and limits¶

Review-only, offline. This audit measures detector skill on one public recording; it is not a clinical sleep-staging product.
One subject, one night. The operating point is chosen on the same recording's Wake null and validated by permutation; generalisation to other subjects, age groups, or sleep disorders requires a separate audit.
Not a universal N3 detector. The honest result is specific to this Fpz-Cz channel, this scoring function, and this recording.
Amplitude, not phase coherence. The detector uses the delta-band Hilbert envelope because a cross-band phase-coherence score did not clear the bar on this corpus. The finding is that slow-wave sleep is amplitude-defined here, not that SPO's phase-coherence machinery is generally unsuited to sleep.

The early-warning matched-false-alarm study (early_warning_matched_false_alarm.md) introduced the honest-evaluation protocol used here.
The ISO-NE forced-oscillation case study (../validation/iso_ne_case1_forced_oscillation.md) shows the same sealed-evidence pattern on a power-system disturbance.
experiments/sleep_staging_eeg.py contains the exploratory cross-band-R heuristic that motivated this audit.