Kahlus-STF Public Dataset Review#

This is a source-selection note for the first public-data STF smoke. It is not a clinical validation claim and does not permit A100 training by itself.

First Target: CHB-MIT Scalp EEG Database#

Source: https://physionet.org/content/chbmit/1.0.0/

Reason for first use:

  • public scalp EEG epilepsy dataset on PhysioNet

  • record-level lists are available through RECORDS and RECORDS-WITH-SEIZURES

  • seizure annotation sidecars use .edf.seizures

  • common benchmark target in seizure detection/prediction papers

Local audit command:

PYTHONPATH=src python3 scripts/fetch_chb_mit_smoke_subset.py \
  --dataset chb_mit_physionet \
  --out-root /tmp/kahlus_chbmit_smoke_subset \
  --patients 2 \
  --records-per-patient 2

The fetch command materializes a small CHB-MIT subset outside the repository. It downloads only selected EDFs, RECORDS, RECORDS-WITH-SEIZURES, patient summary text files, and required .edf.seizures sidecars.

PYTHONPATH=src python3 scripts/run_stf_public_data_audit.py \
  --dataset chb_mit_physionet \
  --data-root /tmp/kahlus_chbmit_smoke_subset \
  --out-dir /tmp/kahlus_stf_chbmit_audit

The audit checks metadata shape only. It does not parse EDF, download data, copy raw signals, or launch A100 jobs.

Local EDF smoke command after the audit passes:

PYTHONPATH=src python3 scripts/run_stf_chb_mit_smoke.py \
  --dataset chb_mit_physionet \
  --data-root /tmp/kahlus_chbmit_smoke_subset \
  --out-dir /tmp/kahlus_stf_chbmit_smoke \
  --max-records 4 \
  --max-samples-per-record 900000 \
  --max-channels 8

The smoke reads capped local EDF records with edfio, preserves variable-length record windows, builds patient-held-out plus time-held-out tasks, and reports persistence, ridge-AR, TinySSM, shuffled-target control, cycle/time-of-day, event-frequency, logistic-ridge, and time-shifted-label rows when CHB-MIT summary text files are present. Binary .edf.seizures parsing remains out of scope; summary text is the event-interval source for this smoke.

Literature Stop#

The research risk is not lack of papers; it is over-optimistic validation. Shafiezadeh et al. report that patient-independent validation is rare in EEG seizure prediction, and Wong et al. emphasize that public EEG datasets differ in structure enough to hurt reproducibility. Therefore, Kahlus-STF starts with patient-held-out and time-held-out audits before any neural architecture upgrade.

Blocked#

  • committing raw EDF or annotation files

  • diagnosis, treatment, medication, stimulation, seizure-prevention, or vEEG/PSG replacement claims

  • A100 runs before local synthetic and public-data smokes pass