Skip to content

Regime-Adaptive Ensemble: An Honest Cross-Corpus Test

Abstract

The CAP variant panel suggested that amplitude-gated sustained coherence beats the delta envelope on a rich montage, and the Sleep-EDF transfer test showed the coherence approach collapses on a sparse two-channel montage. Those two results motivated a label-free regime router: choose the envelope on sparse montages and coherent_sustained_kuramoto on rich ones, using only the observable channel count. This study audits two such routers alongside every component detector on a combined five-recording cross-corpus manifest (four CAP + one Sleep-EDF), at matched false-alarm 0.10 with a 10 000-permutation test.

The result is a negative one, reported as found: the channel-count router does not beat the plain delta envelope. The envelope is the most robust single detector across the five recordings, and the coherence advantage seen on CAP turns out to be recording-specific rather than a clean montage-size effect.

Results

Per-recording detection rate (matched FA 0.10), with the montage router's route in bold:

Recording channels envelope mean-R coherent-sustained montage router
n1 8 0.380 0.231 0.377 0.377 (coherent)
n2 3 0.005 0.223 0.081 0.005 (envelope)
brux2 6 0.913 0.062 0.893 0.893 (coherent)
narco2 3 0.803 0.218 0.856 0.803 (envelope)
SC4001E0 2 0.995 0.000 0.268 0.995 (envelope)

Cross-corpus mean detection rate:

Detector mean DR
normalized_delta_envelope 0.619
regime_adaptive_montage 0.615
regime_adaptive_full 0.615
coherent_sustained_kuramoto 0.495
amplitude_gated_delta_kuramoto 0.421
adaptive_channel_kuramoto 0.163
multi_channel_delta_kuramoto 0.147
snr_weighted_delta_kuramoto 0.140
sustained_delta_kuramoto 0.139

Interpretation

  • The plain delta envelope is the most robust single detector (mean DR 0.619). No detector and neither router exceeds it on the five-recording panel.
  • The channel-count router does not beat the envelope (0.615 vs 0.619). It is bounded below the envelope and above coherent_sustained because it is a weighted mix of their per-recording scores. The router is a strict function of the envelope and coherent-sustained columns above.
  • The routing threshold is too crude, and for a revealing reason. The router treats n2 and narco2 (three bipolar derivations) as sparse and sends them to the envelope. But on narco2 the coherence detector is actually better (0.856 vs 0.803) — so the router leaves 0.053 of detection rate on the table there. Channel count is not a sufficient statistic for the regime.
  • The "montage richness" story from the CAP panel was too simple. coherent_sustained does not uniformly win on rich montages: it loses on brux2 (0.893 vs 0.913) and ties on n1 (0.377 vs 0.380). Its CAP-panel advantage (0.552 vs 0.525) is driven almost entirely by narco2 and by the near-zero-rate n2, not by montage size. The genuine, transferable finding is narrower than first framed: amplitude gating keeps the Kuramoto family alive where plain coherence dies (confirmed on both corpora), but it does not give a robust, montage-indexable advantage over the envelope.
  • The n2 refinement is inert here. regime_adaptive_full is identical to regime_adaptive_montage on this panel: n2 is a three-channel recording, so the sparse-montage rule routes it to the envelope before the low-SNR / high-coherence check can fire. The in-sample n2 axis therefore has no effect and remains unvalidated.

What would be needed to beat the envelope

The evidence points to a regime signal finer than channel count — the recording where coherence helps (narco2) and the recordings where it does not (brux2, n2, Sleep-EDF) are not separated by montage size. A useful router would need a label-free feature that predicts when spatial slow-wave coherence adds information over amplitude alone, validated out-of-sample on more recordings than this five-recording panel provides. Until then the honest recommendation is the plain delta envelope, with amplitude-gated coherence reserved for montages and recordings where it has been shown to help.

Reproduction

PYTHONPATH=.:src python bench/regime_adaptive_ensemble.py \
  examples/real_data/regime_adaptive_ensemble

The script loads the four CAP recordings and the Sleep-EDF recording, audits all seven component detectors plus the two routers at matched FA 0.10 with a 10 000-permutation test, and writes sealed audit records plus the aggregate comparison JSON. The committed evidence is guarded by tests/test_regime_adaptive_ensemble_evidence.py.

Scope and limitations

  • Five recordings. Four CAP + one Sleep-EDF is a small, heterogeneous panel; the cross-corpus means are descriptive, and the negative result is a finding about this panel and this router, not a proof that no router can help.
  • In-sample thresholds. The router thresholds are read from the CAP diagnostic; the n2 axis in particular is unvalidated and, as shown, inert here.
  • Raw EDF files are citation-only. Only derived sealed records and the aggregate JSON are committed.