Post-Ictal Heart Rate Oscillations in Partial Epilepsy

Seven single-lead ECG recordings — 16.77 hours at 200 Hz — made during continuous EEG/ECG/video epilepsy monitoring of five women aged 31 to 48 with partial seizures of frontal or temporal origin and no clinical evidence of cardiac disease, at Boston’s Beth Israel Deaconess Medical Center. It is the smallest dataset in this catalogue.

The database exists to show one finding: transient 0.01–0.10 Hz heart-rate oscillations, two to six minutes long and 15 to 41 bpm peak-to-trough, appearing immediately after a seizure and absent from the pre-ictal period of the same seizures. The paper reports them after 5 of 11 recorded seizures, once in each of the five patients. What makes that analysable is the shipped times.seize file, which gives seizure onset and offset read from the simultaneous EEG by an electroencephalographer blinded to the heart-rate analysis.

All 39 shipped files verify against the release’s own SHA256SUMS.txt, all 7 records pass every ECGBench quality check, and clean/ therefore equals original/.

This is five patients. Use it to study post-ictal dynamics against blinded, EEG-derived seizure times — not to train a classifier. There is no negative class, no control cohort, and after grouping, no more than five independent people in the entire release.

This is the most consequential thing on the page. The release ships no subject identifier anywhere: not in the headers, not in RECORDS, not in the annotations. The paper describes five patients. Reading “there is no column” as “one record per patient” — the natural reading, and what patient_id_column: null would have said — puts three recordings of the same woman on both sides of the split.

sz02, sz03 and sz04 are one woman. The evidence, in the order it was established:

1. Beat morphology. Median beats over a P-through-T window (−0.30 s to +0.50 s about each N annotation, baseline-removed and divided by QRS peak amplitude, so the T-to-QRS ratio survives) correlate at 0.9989 between sz03 and sz04, 0.9844 between sz02 and sz04, and 0.9806 between sz02 and sz03. The control is each record’s own first half against its own second half — the ceiling a same-subject pair can reach — which is 0.9987 for sz03 and 0.9984 for sz04. sz03 and sz04 resemble each other more closely than either resembles its own other half. No other pair comes near: the best is sz01–sz07 at 0.8527 against sz01’s ceiling of 0.9980, and four pairs correlate negatively because their T waves are inverted relative to one another. A beat-level nearest-template assignment agrees — 40.8% of sz04’s beats match sz03’s template better than their own, against 3.7% for sz01 to sz07.
2. The subject count comes out at exactly five, which is the figure the paper states, and it is the check that licenses the grouping rather than merely permitting it.
3. The seizures then fall the way the paper says they do. It reports “Two of the subjects had multiple recorded seizures”. Under this grouping subject 2 has 5 and subject 4 has 2, while subjects 1, 3 and 5 have one each — exactly two. The arithmetic discriminates: of the groupings that give five subjects, only one other also gives two multi-seizure subjects ({sz05, sz06} + {sz02, sz03}), and morphology rejects it outright (sz05 to sz06 correlates at 0.7278 against sz05’s ceiling of 0.9970). Every other five-subject grouping implies three or four multi-seizure subjects and contradicts the paper.

A weaker fourth signal agrees and is independent of the waveforms: the .dat files of sz02, sz03 and sz04 were digitised within five minutes of one another (26 Mar 1998, 17:35 / 17:37 / 17:40) and share a gain of 25 adu/mV.

The values are szdb_subj_1 … szdb_subj_5 — deliberately named so they cannot be mistaken for a PhysioNet identifier — and every row carries subject_id_is_reconstructed: true. Recheck it yourself; it reads all 16.8 h, so it is not run at load time:

from ecgbench.labels.szdb import verify_subject_grouping
verify_subject_grouping("/path/to/szdb/1.0.0").head()

Seven records from five subjects cannot make ten folds, so szdb.yaml sets n_folds: 5 — the only config in ECGBench that changes it. ecgbench splits --dataset szdb picks it up with no flag, which matters because manifest.json records the fold count and hashes the partition into fold_digest: a user who had to remember --n-folds 5 would otherwise compute a different digest and have no way to see why.

Both failure modes are real, and only one of them is loud:

With the default mapping that gives train = folds 1–3 (5 records), val = fold 4 (sz01), test = fold 5 (sz06).

There is nothing to stratify on, and that is measured rather than assumed. Every fold is one subject, and StratifiedGroupKFold raises when every class holds fewer records than there are folds — so over 7 records and 5 folds a usable axis needs a class of 5. Seizure count (1 vs >1) gives 3 subjects against 2 and raises; annotated atrial fibrillation gives 1 against 4 and raises; ST burden per subject is 4, 25, 5, 1, 2 episodes, so no cut reaches 5; record length gives 3 records against 4 and does not raise, but it is not constant within subject 2 (1.5 h, 3.5 h, 3.8 h), and an axis that varies inside a group cannot balance a grouped split. The stratification label is therefore the constant cohort_label, which reduces the split to a plain partition of the five subjects.

Every figure on this page is recomputed from the shipped files. Two disagree with the paper, and both matter.

The seizure count is 10, not 11. Any per-seizure figure derived from this database is a figure over 10, and nothing in the release explains the omission — there is no changelog of any kind here. The paper’s claim that oscillations followed “five of 11 seizures” cannot be reproduced from the shipped files at all, because which seizures they followed was never released either.

The headers declare an ADC resolution of 12 and the data is 8-bit. Values span exactly 256 levels, and four of the seven records (sz03, sz04, sz06, sz07) sit at both rails for hundreds of samples. Because the gain differs by record, the rail in millivolts does too: [−4.0, +6.2] mV at the 25 adu/mV of five records, and [−10.0, +15.5] mV at the 10 adu/mV of sz05 and sz06. ECGBench’s amplitude_range_mv is the union of the two, so no record can fail validation for a rail its own header declares.

Beat, ectopy and HRV figures below are unaudited detector output — the shipped ANNOTATORS file says “unaudited beat annotations from an automated detector”, and the .ari extension names it (ARISTOTLE). Every audited MIT-BIH database in this catalogue uses .atr instead.

73,919 annotations across the 7 records, and three of the symbols are not what a reader expects.

• ? (350, exactly 50 per record) is the detector’s learning phase, not unclassifiable beats. It is always annotations 0–49, covering the first 42–51 s of each record. These are QRS complexes the detector located but had not yet learned to classify, so ECGBench counts them in n_beats and in the RR series and folds them into AAMI Q — which means aami_Q starts at 50 per record. n_learning_beats is exposed separately so the two can be split apart again; genuinely unclassifiable beats number 315 across the release, 278 of them in sz06.
• s (74) is ST change, not a beat. Its aux_note values — (ST0+, ST0+), (ST0-, ST0-) — delimit 37 ST episodes, 31 depression and 6 elevation. Counting s as a beat would inflate every beat total and put a non-beat into the AAMI reduction. Burden is very uneven: sz02 holds 24 depression episodes over 336 s, sz06 a single one lasting 826 s, and sz04 none at all.
• + (2) is rhythm, and it exists in one record only. sz02 carries (AFIB at 10,508.6 s and (N at 10,526.0 s — 17.4 s of atrial fibrillation, ending 25 s before that record’s second seizure begins. The other six records carry no rhythm marker at all, so their af_secs of 0.0 means “never assessed”, not “no atrial fibrillation”; read has_rhythm_annotation first.

There is no signal-quality annotation layer — no ~ and no | anywhere — where mitdb, svdb and nsrdb all carry one. Beat annotation, by contrast, covers essentially the whole of every record: the first beat falls 0.1–0.9 s in and the last within 0.6 s of the end, in all 7.

One unexplained anomaly, recorded so nobody rediscovers it: sz07 alone carries a non-zero num field on 8,768 of its 8,892 annotations, spanning the full signed byte range. It correlates with neither ST deviation at J+60 to J+120 ms (|r| ≤ 0.18) nor the preceding RR interval (r = −0.05), nothing in the release documents it, and the other six records leave num at 0 throughout. ECGBench ignores it.