ECG-ID Database

The reference dataset for ECG biometrics. 310 twenty-second Lead I recordings from 90 volunteers, collected in 2004–2005 by Tatiana Lugovaya at the Electrotechnical University “LETI” in Saint Petersburg for a master’s thesis that asked whether an electrocardiogram identifies the person who produced it. Her answer was 96% correct identification over these 90 individuals.

The label is the person. There is no diagnosis anywhere in this release, no clinical assessment and no rhythm annotation — the ground truth is which of the 90 subjects a recording came from, and the cohort is students, colleagues and friends of the author rather than patients. subject_id is therefore both the label and patient_id_column, which has a consequence spelled out in its own section below: ECGBench’s folds cannot be used for the identification task.

Every record stores the same lead twice. The two channels are ECG I (raw) and ECG I filtered, so the returned tensor is (2, 10000) and both rows are Lead I — one is the author’s offline preprocessing of the other. The thesis deliberately switched off every filter in the cardiograph software, on the grounds that filtering might suppress features useful for identification, so the raw channel carries real baseline drift, 50 Hz interference and high-frequency noise. leads=["ECG I"] is how you avoid handing a model the same lead twice.

Length is uniform, unusually for a PhysioNet release. All 310 records hold exactly 10,000 samples at 500 Hz — 20.000 s, no exceptions — so any window=(start, length) inside (0, 10000) fits every record and WindowOutOfRangeError cannot fire here.

The .atr annotations cover only the first half of each record. Exactly 10 R-peaks and 10 T-peaks per record, from an automatic detector the release states was never audited, all inside the first 5.1–11.7 s. Section below.

Twenty of the 90 subjects were recorded on more than one day, up to six sessions spanning 156 days, which is what makes this database usable for studying whether an ECG biometric persists over time. The other 70 have all their records from a single session.

Every figure on this page was recomputed from the 310 headers and 310 .atr annotation files, after verifying the shipped copy against the release’s own SHA256SUMS.txt — all 977 listed data and documentation files match. The 22 listed files absent from a normal download are website assets under .old/ (biometric.shtml and 21 PNGs), not data.

Three of the release’s own headline figures reproduce exactly and one does not. The README and the PhysioNet abstract give 310 records, 90 subjects, 44 men and 46 women, ages 13 to 75 — all four confirmed. Both then say the number of records per subject runs “from 2 (collected during one day) to 20 (collected periodically during 6 months)”.

RECORDS and the files on disk agree with each other and disagree with the prose. There is no changelog in the release to explain it, and the directory timestamps show the 90 subject directories were written in two batches (September 2011 and April 2012) while the README was last touched in April 2012 — so the likeliest reading is that the prose describes the collection protocol rather than the final release. ECGBench reports the shipped figures.

Age is a subject attribute here, not a per-record one. No subject’s recorded age changes across sessions, even the two spanning 140 and 156 days: Person_02 is 23 in all six of its sessions. So a per-record age is safe to cross with sex for fold construction.

There are only ten acquisition days in the whole release, and one holds 43% of it. ecg_date is a session identifier as much as a date, which is why the session columns are derived from it.

Any equipment, electrode or environmental peculiarity of 12 May 2005 is present in 134 of the 310 records.

This is the most important thing on this page.

ECG-ID’s ground truth is identity, so its label column is its patient column. ECGBench groups folds by patient, which means every subject’s records sit wholly inside one fold and no fold’s model has ever seen the person it would be asked to recognise. Folds hold 7–11 subjects each and no subject spans a fold.

That is the right default for every other use of these recordings. A model trained on 21 of Person_02’s 22 records and evaluated on the 22nd is measuring nothing; leaving the grouping off would produce exactly that, for the 89 subjects with more than one record. It is the wrong default for the one task this database was built for.

For identification, split within subject. The label loader exposes the session structure for precisely this:

ds  = ECGDataset("ecgiddb", split="train", labels=True, data_path=...)
ids = ds.metadata_df["record_id"]
multi   = ds.labels_df["is_multi_session"]          # 20 subjects across the release
enrol   = ids[multi & (ds.labels_df["session_index"] == 1)]
verify  = ids[multi & (ds.labels_df["session_index"] >  1)]

Holding out a subject’s later sessions is the protocol the thesis’s persistence experiment used, and it is the harder and more meaningful one: the electrodes were reattached from scratch for every recording, so a same-session pair shares an electrode placement that a cross-session pair does not.

The longest-running subjects are Person_02 (22 records, 6 sessions, 156 days) and Person_01 (20 records, 5 sessions, 140 days). Together they are 42 of the 310 records.

The thesis’s own 195/115 train/test division is not recoverable. It divided these same 310 records into 195 training and 115 test records, chosen for “maximum difference between records in different sets both in monitoring time and human physical state” — and that assignment appears in no file. Not in RECORDS, not in a header comment, not in a separate list. So has_predefined_splits is false for lack of the data, not for lack of a split, and no ECGBench figure is comparable to the thesis’s 96%.

Both channels of every record are Lead I, the left-hand minus right-hand potential, taken with limb clamp electrodes with the subject seated. The thesis chose Lead I because it is easy to acquire and insensitive to small electrode displacements, imitating “the likely scenarios of user interaction with a practical identification system”.

The filter chain, from the thesis: baseline drift removed by level-9 db8 wavelet decomposition with the final approximation subtracted, an adaptive bandstop at 50 Hz for power-line noise, and a 5th-order Butterworth lowpass (Wp 40 Hz, Ws 60 Hz, Rp 0.1 dB, Rs 30 dB).

It is zero-phase, verified rather than assumed: peak cross-correlation between the two channels is at lag 0 in 273 of the 310 records and at ±1 sample in the other 37. So the channels are sample-aligned and the .atr annotation indices apply to both.

Because the residual raw − filtered is therefore noise rather than a phase difference, it is a usable per-record noise level. ecgbench.labels.ecgiddb.scan_noise_levels(data_path) computes it — it is a separate call rather than a label because it needs the waveforms decoded:

Correlation between the two channels runs 0.103 to 0.978 (median 0.691). A low value means drift dominates the raw channel, not that the filter failed.

config.leads is 2 because that is the shape of the returned tensor; this page’s leads field says 1 because that is the number of electrode pairs. Both are true of different things, and a test pins the pair so neither gets “corrected” to match the other.

The shipped ANNOTATORS file is one line: “atr — unaudited R- and T-wave peaks annotations from an automated detector”. Taking it at its word:

• Exactly 20 annotations in every one of the 310 files — 10 N R-peaks and 10 t T-peaks, strictly alternating N t N t …. No record has more, none fewer, and no other symbol appears anywhere in the release.
• They describe the first ten beats and nothing after them. The last annotation lands at sample 2,542 to 5,869 of 10,000 (mean 4,002, median 3,970) — 5.1 s to 11.7 s into a 20.000 s record. Only 18 records have any annotation past the halfway mark, and 8.3 s to 14.9 s of every record is unannotated.
• annotated_fraction carries this per record, 25.4% to 58.7%. A window past last_annotation_sample contains beats nobody marked.

N here means “the detector found a beat”, not “normal beat”: no beat in this database was ever classified. So mean_hr_bpm (50.0 – 132.7, median 76.7), sdnn_ms (4.1 – 393.5) and mean_rt_interval_ms (169.8 – 452.2) are estimates over nine RR intervals from an unaudited detector. They describe the record; they are not an HRV result, and this is not a beat-detection reference — use qtdb or ludb for that. mean_rt_interval_ms is R-peak to T-peak, not a QT interval, because the annotation marks the T peak rather than its offset.

The wide SDNN range is real and expected: the thesis deliberately did not restrict heart rate or physical and emotional state, so within-record rate variation is part of the design rather than a data-quality problem.

No record has a NaN sample, a flat or all-zero channel, an unreadable header or a truncated signal. The only check that fires is amplitude_outlier, on three records:

The bound is [-10.24, 10.235] mV, and it is a physiological bound that merely looks like a hardware rail. Every one of the 620 signal lines in the release reads 16 200 12 0 — format 16, gain 200 adu/mV, adc_res 12, adc_zero 0, no baseline — which nominally confines every sample to [−2048, 2047] adu. The stored int16 samples go far past it: Person_76/rec_2 reaches −30,831 adu. The 12 bits describe the cardiograph’s converter, and the baseline drift the thesis chose not to filter rides on top of it. So unlike apnea_ecg, which uses the same two numbers as a rail its data sits inside, here the data is not rail-bounded at all.

No float32 slack is needed, and that was checked rather than assumed: the closest passing record reaches 10.115 mV and −3.060 mV, nowhere near the bound, so the rail-rounding trap that bit chfdb cannot arise. Widening the bound would only readmit two records whose drift is fifteen times the physiological range.

Person_47/rec_2 is a different case from the other two — a genuine 10.885 mV excursion rather than gross drift — and it is excluded on the same rule. version="original" returns all 310 with quality_issues attached.

Folds are built with StratifiedGroupKFold, grouped on subject_id and stratified on stratify_class — sex crossed with a single age cut at 30 years.

Why that cross and nothing finer. A class has to contain at least as many subjects as there are folds to appear in every fold, and the smallest cell here holds exactly 10 — the floor for 10 folds. Crossing sex with four age bands instead produces 5-subject cells and leaves empty cells in the fold table. Neither axis works alone either: stratifying on sex alone leaves four folds with no subject over 45 (age is skewed — 45 of the 90 subjects are 21–30 and only 11 are over 45), and stratifying on age bands alone leaves one fold at 9 female / 19 male records, because group sizes run 1 to 22 and a single subject can swing a fold.

Heart rate was rejected as an axis: it is not a subject attribute here, and it comes from an unaudited detector over ten beats. stratify_class is a fold-construction device and must not be trained on as one.

Use the folds, not the default split. With folds 1–8 → train, 9 → val, 10 → test:

Seven subjects is not an evaluation set. For a real evaluation, cross-validate: split=None with fold_numbers=[...] selects by fold from folds.csv and ignores the default layout.

Folds 4 and 10 are the largest because they drew Person_01 (20 records) and Person_02 (22). Sex is balanced 13–24 female against 13–25 male per fold; those two subjects are what pushes the extremes.

No related: edge is declared. Two other releases in this catalogue come from Russia — ludb (Lobachevsky University, Nizhny Novgorod) and incartdb (St. Petersburg Institute of Cardiological Technics) — and neither shares a recording, an institution or a decade with this one: ECG-ID is a 2004–2005 LETI student cohort with one lead at 20 s, ludb is 200 twelve-lead hospital records at 10 s, incartdb is 75 twelve-lead 30-minute Holter excerpts. Nothing else in the catalogue is a biometric-identification release.

The overlap that matters is inside this release — 89 of the 90 subjects contribute more than one record — and it is handled by the subject grouping.