tOLIet (Thigh-based ECG, toilet seat)

An ECG taken through the backs of your thighs while you sit down. The electrodes are dry polymer pads moulded into a toilet seat: no gel, no skin preparation, no operator placing anything, and no cooperation from the subject beyond sitting. 145 sittings by 86 volunteers at the Centro Hospitalar Universitário de Lisboa Central, published on PhysioNet in February 2026.

The release is a controlled comparison of electrode texture. Four electrode pairs are embedded in the same seat, differing only in the surface moulded into them — flat, sinusoidal, pyramidal, trapezoidal — and all four record the same thigh-to-thigh derivation at the same instant. That is the scientific variable, and the answer it gives is blunt: see the texture table below.

One ECGBench record is one electrode channel, not one sitting. 145 files × 4 electrodes = 580 records, each a single-lead (1, samples) tensor named <sitting>_<channel> — 15_1_A2 is the sinusoidal channel of subject 15’s second sitting. The reason is in the next section, and it is not cosmetic.

238 of those 580 channels are electrodes that never made contact, and they are exactly what separates original from clean. The default version="clean" gives you the 342 channels that recorded something.

Nothing here is a diagnosis dataset. Four of the 145 sittings carry a free-text paroxysmal atrial fibrillation note and the other 141 carry nothing. The stated uses are electrode-texture comparison and biometric identification; for the latter the label is the subject, which is also the grouping column, so ECGBench’s folds cannot serve it — same caveat as ecg-id-database.

The files look like four-lead records: each .txt holds columns A1–A4 sampled together. Representing them that way is what ECGBench does everywhere else, and here it produces a dataset nobody can use.

An electrode pair that made no contact reads a constant ADC code for the whole sitting. Kept as a four-lead record, that is a flat lead inside an otherwise good record — and flat_line rejects the whole record when any lead fails. Only 5 of the 145 sittings have all four electrodes live, so a four-lead record set gives a clean version of 5 records and two empty folds. Split per channel and clean is 342 real single-lead ECGs.

The cost is that 580 records are not 580 independent observations. The four channels of a sitting are the same beats seen by four sensors, so the independent unit is the sitting (145) or the subject (86). Folds group by subject, so nothing leaks across a fold boundary — but a model reported as having been evaluated on 580 samples is overstating its evidence. Group on source_record before counting.

Every figure on this page was recomputed from the shipped files, after verifying the local copy against the release’s own SHA256SUMS.txt — all 174 listed files match.

The release says 149 recordings and ships 145. DataSet.csv lists 149 IDs; ECG_EXP/ holds 145 .txt files. 12_1, 13_1, 14_1 and 41_1 are tabulated and absent from the download, with no changelog in the release to explain it. ECGBench drops those four rows with a warning rather than emitting records that would all fail corrupt_header. The subject count is unaffected, because each of the four is a later sitting of a subject whose first sitting is present.

The published demographic means are per subject, not per recording — that is how they reproduce to the second decimal. The per-sitting age mean is 29.99 ± 10.59, so quoting “mean age 31.7” alongside a record count mixes two denominators.

records: 580 on this page is ECGBench’s record count, not the release’s. The release has 145 recordings; ECGBench exposes each of their four electrode channels separately, for the reason in the section above.

This is the trap on this dataset, and it survives validation.

The front end is ±1.5 mV full scale into a 10-bit converter, so every sample is inside the configured amplitude_range_mv by construction and amplitude_outlier cannot fire on a single record. What actually goes wrong is saturation at the rail — poor contact drives the amplifier into one end and keeps it there.

A channel pinned at one rail has a tiny variance and flat_line catches it. A channel oscillating between both rails has a large variance, passes flat_line, and is not an ECG:

12 of the 342 active records are at a rail for more than half their samples, and 130 touch one at all — over 66 of the 145 sittings. Ten of the twelve worst are the trapezoidal electrode. No check in CHECK_REGISTRY measures clipping, so ECGBench does not exclude them; clipped_fraction, min_mv and max_mv are in the labels so you can:

usable = ds.labels_df[ds.labels_df["clipped_fraction"] < 0.01]

The highest ADC code occurring anywhere in the release is 1022, not 1023.

Records run 14,400 to 197,250 samples at 1 kHz — 14.4 s to 3 min 17 s, median 126.3 s — against the landing page’s “up to 5 minutes per session”. expected_samples is therefore deliberately empty in the config, which is the documented escape hatch for genuinely variable-length data.

window=(0, 14400) is the largest fixed window that fits every record; one sample more raises WindowOutOfRangeError on record 79. The window is pushed into the reader’s skiprows/nrows, so on a 197-second record it decodes 14 seconds rather than decoding everything and slicing.

The signals ship as PLUX/BITalino OpenSignals text exports: three # preamble lines, the second a JSON blob naming all eleven columns and their bit depths, then tab-separated integers. ECGBench gained a opensignals reader for this release. The signal path names the column it wants:

ECG_EXP/15_1.txt:A2

because the same rows also carry a sequence number, four digital I/O columns and two 6-bit analog channels that are zero throughout.

The reader returns fractions of full scale, not millivolts, and the config supplies signal_unit_scale: -3.0. That is not a unit conversion in the usual sense — it is the amplifier’s full-scale span in millivolts, and it is negative because the seat’s differential front end inverts. Together they reproduce the release’s own Script/read_ecg_data.py

((1024 - raw) / 1024 - 0.5) * (33 / 11)

bit for bit; that equality was checked sample-for-sample against a raw np.loadtxt of the files rather than assumed. Code 0 is exactly +1.5 mV, which is why an unconnected electrode reads a flat +1.5 mV line rather than a zero line — and why missing_leads never sees one but flat_line does.

ECG_REF/ holds 23 Sapphire-format .XML files — 10-second 12-lead resting ECGs from a hospital cardiograph, in microvolts — recorded alongside the seat recordings. They cover 23 of the 86 subjects (58, 59, 60, 67, and 68–86), and every one names a sitting that exists in ECG_EXP/.

ECGBench does not load them and gives them no records. A 10-second clinical 12-lead ECG is a different modality from a two-minute thigh recording; putting both in one record set would mean folds containing both, and a leads= argument that means two different things depending on the row. has_reference_ecg (true for 92 of the 580 records — the four channels of each of the 23 sittings) and reference_path point at the files, and the release’s own Script/read_ref_data.py parses them.

This is the ground truth the abstract refers to, so if you are validating thigh-derived morphology against a clinical lead, those 23 sittings are the whole of the available evidence.

flat_line is the only check that fails anything: 238 records, one issue each. No record has a NaN sample, an unreadable header or a truncated signal, and amplitude_outlier cannot fire at all (see above). missing_leads finds nothing because a dead electrode reads +1.5 mV rather than 0.

ECGBench does not invent a threshold for “did this electrode record anything”: ecgbench.labels.tollet runs the project’s own check_flat_line per channel and exposes the verdict as signal_active, so the label column and the validation report are the same decision and clean is exactly the signal_active records.

Folds are built with StratifiedGroupKFold, grouped on subject_id and stratified on stratify_class — sex crossed with signal_active.

Why the cross. Liveness is in it because it decides how big a fold is in the version most people load. Measured over the shipped files at random_state=42:

Electrode texture is deliberately not in the cross and does not need to be: every sitting contributes all four textures, so any partition of sittings splits them evenly by construction — 13 to 15 active A1 channels per fold. Putting it in explicitly would fail anyway, because A3_active has 7 records from 7 subjects and a class needs at least ten subjects to appear in ten folds.

No subject and no sitting spans a fold.

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

Nine 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.

No related: edge is declared. This is a 2020s Lisbon cohort recorded on purpose-built hardware that exists nowhere else in the catalogue, and no other release contains a thigh derivation or a dry-electrode seat recording. The nearest neighbour by purpose is ecg-id-database, which shares the biometric-identification framing and the “the label is the subject” problem, but not a single recording, subject or institution.

The overlap that matters is inside this release, twice over: 53 of the 86 subjects contributed more than one sitting, and every sitting contributes four records of the same beats. Both are handled by grouping folds on subject_id.