Apnea-ECG Database

The reference dataset for detecting sleep apnea from the ECG alone, and the data of the PhysioNet/CinC Challenge 2000. 70 overnight single-lead recordings — 573.7 hours at 100 Hz — with an expert apnea annotation for every one of their 34,313 minutes, of which 13,064 are apnea.

The ground truth is per minute, not per record. Each .apn file holds one A/N annotation per minute of recording, assigned by an expert from simultaneously recorded respiration and oxygen saturation. ECGBench exposes the whole sequence as apnea_sequence, so minute i is labelled by apnea_sequence[i] and window=(i * 6000, 6000) returns exactly that minute. The record-level apnea_class is a whole-night summary the challenge used to describe records; the task is the minutes.

The 70 records come from 30 subjects, and the release’s own learning/test split leaks 18 of them. This is the most important thing on this page. Apnea-ECG publishes no subject identifier anywhere — not in the headers, not in RECORDS, not in the annotations — so nothing warns a user that 27 of the 30 subjects contributed between two and four nights. 18 of those subjects, 49 of the 70 records, have recordings on both sides of the challenge’s a/b/c vs x division. ECGBench therefore does not adopt that division: patient_id_column is a reconstructed subject_id and folds are grouped on it. challenge_set survives as a label so the original 2000 result stays reproducible, with the leak stated rather than inherited.

Two pairs of records are the same recording, bit for bit. x35 is x22 shifted by 40 s and c06 is c05 shifted by 80 s. Both are kept — each is an official record with its own official annotations — and the grouping puts each pair in one fold. Section below.

Records are whole nights: 2,430,000 to 3,462,000 samples. Batching needs a window=(start, length), which is read at load time rather than cropped afterwards. Length is not uniform — 6.75 h (x17) to 9.62 h (a12) — so a window sized for one record need not fit another.

The single channel is not a named lead. All 70 headers call it ECG and the release documents no electrode placement anywhere, so leads=["ECG"] selects a channel position. Do not stack it with a 12-lead dataset’s lead I or II by name.

Every figure on this page was recomputed from the 70 headers, the 70 .apn and .qrs annotation files and additional-information.txt, after verifying the shipped data against the release’s own SHA256SUMS.txt — all 383 listed data files match. The only listed file absent from this copy is annotations.shtml, a documentation page.

The 35 x*.apn files are deliberately not in that checksum list. PhysioNet withheld the test set’s answers for the challenge and posted them on 2020-06-01, a year after the 2019 checksum file was written. They are genuine, and they are what makes all 70 records usable rather than 35.

Subjects: 30, and the database is usually described as having 32. The four demographic fields published per record (age, sex, height, weight) take exactly 32 distinct values across the 70 records, which is the figure the literature quotes. Two of those 32 groups then merge, because two pairs of records turn out to be the same recording — see the section below — leaving 30. PhysioNet’s own landing page states neither number; it says only “70 records”.

Class counts are derived, and they reproduce the release exactly. The release encodes the class in the record name for the learning set only (a01–a20 apnea, b01–b05 borderline, c01–c10 control) and the withheld test records are all named xNN. Applying PhysioNet’s stated criterion to the .apn counts — control under 5 apnea minutes, apnea 100 or more, borderline between — reproduces all 35 learning-set letters with no exceptions, which is what licenses applying it to the x records. It lands them on 20 A / 5 B / 10 C: the learning set’s exact composition, which is evidently how the test set was built.

apnea_class is a whole-night summary, not the stratification target and not the task. It is a property of the night, not the subject: three subjects contributed nights that fall in different classes (b02 b03 x16 x21 spans A and B, a17 x12 spans A and B, b04 c08 spans B and C). Folds are balanced on it because it is the release’s own taxonomy and 40/10/20 survives 10 folds; the AHI severity bins (23 normal / 5 mild / 11 moderate / 31 severe) do not, since a class of 5 cannot be spread over 10 folds.

The minute totals differ slightly from the published table, in two ways. additional-information.txt scores 34,428 minutes and 13,066 apnea minutes; the annotation files hold 34,313 and 13,064.

c07 holds 429 annotated minutes (7.14 h of signal) against the 454 scored, and c08 513 (8.57 h) against 535. The apnea counts for both still agree exactly (4 and 0), so nothing labelled apnea is missing — but a per-record minute total taken from the paper will not match the file. Both figures are exposed side by side (n_annotated_minutes and published_minutes) so the discrepancy stays visible rather than being silently resolved.

Cohort. Ages 27–63, BMI 19.2–45.3, and 25 of the 30 subjects are men (57 of the 70 records). Mean heart rate runs 48.6–83.9 bpm and SDNN 43.7–227.1 ms over 2,270,442 machine-detected beats.

Apnea-ECG divides itself into a 35-record learning set (a01–a20, b01–b05, c01–c10) and 35 test records (x01–x35), and training on the first while evaluating on the second is the standard protocol in this literature. It puts the same subjects on both sides.

There is no subject identifier in the release to notice this with. What additional-information.txt does publish per record is age, sex, height and weight, and those four fields take exactly 32 distinct values over the 70 records — the subject count the database is described by. Records sharing all four are one subject; two verified duplicate recordings merge two further groups, giving 30.

So 70% of the test records belong to a subject the model has already seen during training. ECGBench sets has_predefined_splits: false and groups its own folds on subject_id, which routes the split through StratifiedGroupKFold; no subject spans a fold. challenge_set is kept as a label column for anyone reproducing a 2000-era result, and the caveat belongs with the comparison.

Over-grouping was preferred to under-grouping throughout. Two genuinely distinct people sharing all four coarse attributes would be merged, which costs a little fold granularity; a missed subject leaks, and nothing downstream can detect it.

x35 is x22, and c06 is c05 — the same night released twice under different names, with slightly different crop windows.

The comparison is on the integer ADC values, so no floating-point tolerance is involved. They were found by counting exact 8-grams of RR intervals shared between every pair of records in the release: these two pairs share 4,152 and 4,006 against a background of 20–90 for ordinary pairs, two orders of magnitude above the noise. Confirmation came from maximum-over-lag cross-correlation, 0.9998 and 0.9629, against 0.003 – 0.054 for every control pair — including same-subject different-night pairs, which is the control that makes the result mean something.

The demographics contradict each other across x22/x35. The published table gives x22 as 27 F, 158 cm, 53 kg and x35 as 31 M, 184 cm, 74 kg, for what the waveform proves is one recording. One of those rows is wrong and there is no way to tell which, so the grouping unions transitively: {x17, x22} and {c01, x35} become one subject. c05/c06 differ less dramatically (169 cm / 57 kg against 171 cm / 65 kg) and merge the same way.

Both records of each pair are kept. Each is an official record with its own official annotations, and dropping one would silently diverge from the release. duplicate_of names the relationship, and because the pair shares a subject_id the split cannot place identical waveform in train and test — which an ungrouped split over these 70 records otherwise would.

All 70 records pass every check, so original and clean hold the same 70 rows. There are no NaN samples, no flat or all-zero channels and no unreadable header anywhere.

• truncated_signal is disabled, by leaving expected_samples empty. Records run 2,430,000 (x17) to 3,462,000 samples (a12) and every one is a complete night, so any single threshold would drop sound records as truncated. Omitting the rate disables the check rather than making it fire — the same escape hatch ptbdb, afdb and nsrdb use.

• amplitude_range_mv is [-10.24, 10.235], the 12-bit rail computed from the hardware: baseline 0 at the declared gain of 200 adu/mV puts every possible sample in [−2048, 2047] adu. Unlike NSRDB, this release actually reaches it — 27 records touch the lower rail and 38 the upper, which is what a night of ambulatory Holter with electrode movement looks like.

  A bound sitting exactly on an attained value is the case where float32 rounding decides whether a check fires, so the direction was checked rather than assumed: float32(10.235) is 10.234999656 and float32(-10.24) is −10.239999771, so both rails round toward zero and a railed sample compares strictly inside the bound. No slack is needed. (Contrast chfdb, whose 10.585 rail rounds outward to 10.585000038 and excluded the very record it was computed from until the bound was widened to 10.586.) What this check guards here is a mis-scaled copy — microvolts, or a re-release with a different gain — which would exceed the range by orders of magnitude on the first record.

RECORDS lists 86 names. The partition is 70, and the other 16 are not additional recordings:

The *er headers point at a01.dat and friends, so their ECG is not merely equivalent — it is the same bytes. Including them would put one recording in the partition twice. The *r records hold no ECG and could not be loaded as ECG at all. A record is kept only when two independent filters agree — its header declares exactly one signal named ECG, and its name matches [abcx]NN — because either alone would silently admit the wrong set if a re-release changed the naming or the channel layout; disagreement is an error rather than something resolved by preferring one filter.

has_respiration flags the 8 ECG records whose companions exist, so a user wanting chest, abdominal and nasal respiration and SpO2 knows which to open directly with wfdb. It is derived from the companions actually present rather than from a constant, so a partial download says so.

The .qrs beat annotations are machine-generated and unaudited. PhysioNet states they came from sqrs125 at per-record thresholds and that “in no case were the annotations hand-edited”. mean_hr_bpm, sdnn_ms and rmssd_ms are computed from them, over RR intervals in [0.3 s, 2.0 s], and describe a whole night of sleep — they are not an HRV result, and .qrs is not a beat-detection reference. The | markers in those files are QRS-like artifacts (5,346 of them) and never enter an RR interval.

Worth stating because the release itself got this wrong first. When the annotations were originally posted, PhysioNet described them as “an ‘A’ annotation indicates that apnea occurs during the following one-minute interval”. It published a correction, and annotations.html now reads:

Each “A” annotation indicates that apnea was in progress at the beginning of the associated minute; each “N” annotation indicates that apnea was not in progress at the beginning of the associated minute.

The first annotation sits at second 0 and describes the interval 0–59.99 s, the second at second 60, and so on. Some papers still quote the withdrawn description. Annotation positions are checked against the exact one-minute grid on load, because apnea_sequence is indexed by minute downstream and a gap would shift every label after it with no error raised.

Folds are built with StratifiedGroupKFold, grouped on the reconstructed subject_id and stratified on apnea_class (40 A / 10 B / 20 C). No subject spans a fold, and both duplicate pairs land in one fold each.

Use the folds, not the default split. ECGBench’s convention is folds 1–8 → train, 9 → val, 10 → test, and with 70 records grouped into 30 subjects that gives train 58 records, val 6 and test 6:

Six records is not an evaluation set, and the class composition of val and test is lumpy by arithmetic rather than by defect: 30 groups over 10 folds is three groups per fold, and a group can hold four records of one class. The minute-level positive rate varies with it, 34.8% to 63.7%. For a real evaluation, cross-validate: split=None with fold_numbers=[...] selects by fold from folds.csv and ignores the default layout. Ten-fold cross-validation over the whole 70 records is what this dataset’s size supports.

Fold sizes are 6–8 records (folds 1–3 hold 8, folds 4, 9 and 10 hold 6).

No related: edge is declared. The only other sleep-apnea release in this catalogue, the St. Vincent’s / UCD Sleep Apnea Database (ucddb), is a separate 25-subject cohort recorded in Dublin with three ECG channels at 128 Hz in EDF — a different institution, a different decade and no shared recordings. There is no catalogue entry for the PhysioNet/CinC Challenge 2000, because Apnea-ECG is that challenge’s data rather than a bundle containing it.

The overlap that does exist is inside this release, and it is handled by the subject grouping rather than by a related: edge: 27 of 30 subjects contribute multiple nights, and two pairs of records are the same recording.