Skip to content

Real-Data Case Study: CHB-MIT Multi-Channel Kuramoto Pre-Ictal Audit

Abstract

We compare the textbook multi-channel Kuramoto order parameter with three adaptive or semi-adaptive variants on real annotated seizures from the CHB-MIT Scalp EEG Database (subject chb01). The task is to separate the 5-minute pre-ictal window from seizure-free interictal epochs at a matched 10 % false alarm rate. A new global top-k PLV detector (mean-R over the 15 channels with the highest mean phase-locking value to the mean field) is now the strongest performer (mean detection rate 0.81, AUC 0.94), beating both the simple mean-R baseline (mean detection rate 0.70, AUC 0.91) and the earlier adaptive weightings. The original SNR+kurtosis adaptive variant lags (mean detection rate 0.43, AUC 0.80), and the PLV-to-mean-field soft weighting is a partial improvement (mean detection rate 0.60, AUC 0.84). The delta-band variants remain uninformative. The result shows that the right form of adaptive channel selection can improve on the textbook mean-R detector for this corpus, but only when the selection is driven by cross-channel phase coherence and restricted to a top-k subset rather than soft-weighting all channels.

This within-subject advantage does not generalise. A leave-one-subject-out audit across chb01–chb05 (see Cross-subject generalisation below), which calibrates k only on the training subjects, finds top-k PLV beats mean-R on just 1 / 5 held-out subjects, with both detectors averaging ≈ 0.50 AUC (chance). The pre-ictal coherence signal is present only on chb01 and, weakly, chb05. The chb01 headline is therefore subject-specific, not a general seizure detector; the fixed-k advantage was an artefact of tuning k on the evaluation subject.

Artefacts

All sealed records, per-seizure summaries, and the aggregate comparison JSON are committed under:

examples/real_data/chb01_seizures_multichannel_kuramoto/

Run the audit yourself:

python bench/chbmit_multichannel_kuramoto.py DATA \
    examples/real_data/chb01_seizures_multichannel_kuramoto

where DATA/ contains the raw chb01_*.edf files from PhysioNet.

Headline results

Detector Band Mean DR @ 10 % FA Mean AUC Geo-mean p Beats chance
mean_kuramoto 0.5–4 Hz 0.016 0.379 0.9314 0 / 7
adaptive_kuramoto 0.5–4 Hz 0.063 0.402 0.7052 0 / 7
mean_kuramoto 4–30 Hz 0.698 0.908 0.0008 6 / 7
adaptive_kuramoto 4–30 Hz 0.429 0.797 0.0118 4 / 7
plv_kuramoto 4–30 Hz 0.603 0.840 0.0035 4 / 7
topk15_plv_kuramoto 4–30 Hz 0.810 0.940 0.0003 7 / 7

Take-away

For seizure-EEG early-warning on chb01, the top-k PLV global selection detector is now the preferred multi-channel Kuramoto variant. Selecting the 15 channels with the highest mean PLV to the mean field and computing the unweighted mean-R over that subset improves both detection rate and AUC over the textbook mean-R detector. The SNR+kurtosis and soft PLV weightings are inferior on this corpus; the gain comes from hard channel selection driven by cross-channel phase coherence, not from continuous quality weighting. This advantage is specific to chb01 and does not transfer — see the cross-subject audit below.

Cross-subject generalisation (leave-one-subject-out)

The chb01 headline tunes k on the same subject it scores. To test whether top-k PLV is a real detector rather than a per-subject artefact, we ran a leave-one-subject-out audit on five subjects. For each held-out subject, k is calibrated only on the other four subjects (the k maximising their mean top-k PLV AUC), then top-k PLV and the mean-R baseline are scored on the held-out subject's pre-ictal windows versus its own interictal nulls. No subject influences its own k, so each figure is a genuine out-of-sample estimate.

Held-out Pre-ictal epochs Calibrated k top-k PLV AUC mean-R AUC
chb01 63 23 0.908 0.908
chb02 18 15 0.114 0.275
chb03 63 23 0.309 0.309
chb04 27 23 0.460 0.348
chb05 45 23 0.707 0.707
mean 0.499 0.509

Out-of-sample, top-k PLV beats mean-R on 1 / 5 subjects — chb04, where both detectors are below chance, so that "win" is between two failing detectors. For four of five subjects the calibrated k is 23, which is all channels, i.e. exactly the mean-R baseline: the training subjects do not support a beneficial k < 23. The one time it does pick k = 15 (chb02), it makes the held-out result worse (0.11 vs 0.28). Mean out-of-sample AUC is ≈ 0.50 for both detectors — indistinguishable from chance.

The deeper finding is about the signal, not the detector: pre-ictal Kuramoto coherence rises only on chb01 (AUC 0.91) and, weakly, chb05 (0.71); on chb02/chb03/chb04 it is at or below chance. Neither mean-R nor top-k PLV is a viable subject-independent seizure predictor on this corpus. The chb01 top-k advantage was k-tuning on the evaluation subject, not generalisation.

Reproduce:

python bench/chbmit_crosssubject_validation.py DATA \
    examples/real_data/chbmit_crosssubject_kuramoto \
    chb01 chb02 chb03 chb04 chb05

where DATA/ holds each subject's chbNN-summary.txt and chbNN_*.edf. Sealed under examples/real_data/chbmit_crosssubject_kuramoto/.

Limitations

The single-subject chb01 headline is real within chb01 but is not a general seizure detector, as the cross-subject audit above establishes:

  • Does not generalise across subjects. Out-of-sample, top-k PLV matches or loses to mean-R on 4 / 5 subjects and both average ≈ 0.50 AUC (chance).
  • The fixed k advantage was tuning. The winning k = 15 was selected on the same chb01 seizures it scored; under honest per-subject k calibration the advantage disappears (calibrated k collapses to 23 = mean-R).
  • The pre-ictal coherence signal itself is subject-specific. It is strong on chb01, moderate on chb05, and absent (≤ chance) on chb02/chb03/chb04.

Any use of this detector must be per-patient calibrated, not deployed as a subject-independent model. Cross-corpus validation (e.g. the CAP sleep database) would further test whether even per-patient calibration transfers across domains.