STAFF III Database

520 recordings from 104 patients undergoing prolonged elective percutaneous transluminal coronary angioplasty (PTCA) at Charleston Area Medical Center, West Virginia, in 1995–96. Almost everything else in this catalogue labels a recording with a diagnosis; STAFF III labels it with a position in a procedure, and that is what makes it valuable: the ischaemia is deliberate, timed, and each patient serves as their own control.

Each patient contributed a short series around one angioplasty:

The 142 inflation records carry 152 balloon inflations — nine records hold two or three — annotated in WFDB .event files with sample-accurate inflation, deflation and contrast-injection instants. Occlusions ran 28 s to 595 s (median 289 s), 10.4 hours of controlled coronary occlusion in total, and the occluded vessel is recorded for every one.

There are 9 signals per record, not 12, and the precordials come first. Files store V1 V2 V3 V4 V5 V6 I II III. aVR, aVL and aVF are exact linear combinations of I and II and were not stored, so the montage is a standard 12-lead one in the clinical sense but signal[0] is V1, not lead I. The declared order was verified against all 520 headers, and against the signals themselves: Einthoven’s III = II − I holds to 2–6 quantisation steps (LSB = 0.625 µV), so the three limb channels are what they claim to be.

Folds are grouped by patient, and that is not optional here. 104 patients contributed 520 records — a mean of 5 each, minimum 2, maximum 7 — and a patient’s baseline, occlusion and recovery are the same heart under the same electrodes minutes apart. Split those across train and test and the score measures patient identity. No patient spans a fold.

Record length leaks the label. Inflation records have a median duration of 518 s against 300 s for every other phase, so a model handed raw lengths can read the phase straight off. Window to a fixed length before training. The shortest record is 94.514 s, so window=(0, 90000) is the largest round window that loads every record.

Every figure on this page was recomputed from the 520 headers, the 142 .event files and the shipped annotation spreadsheet, after verifying all 1,189 files against the release’s own SHA256SUMS.txt (all match). Points where the shipped data differs from what is commonly quoted:

• 104 patients, not 108. File numbers run to 108 but 28, 67, 78 and 103 are unused. The spreadsheet says so itself: “the database contains 104 patients (not 108 as mentioned in some publications since 4 file numbers are unused)”.
• 520 records. This page previously recorded “152 inflations”, which is the inflation count, not the record count — 152 inflations occur across 142 of the 520 records. Both figures are given above.
• 9 stored leads, not 12. Descriptions of STAFF III as “standard 12-lead” are clinically correct — the missing three are derivable — but a program reading the files gets 9 channels. This page’s leads field now says 9.
• The spreadsheet’s D2 field is unreliable and ECGBench ignores it. D0;D1;D2 is time-to-inflation, inflation duration, and time from deflation to end of file. D0 and D1 agree with the .event markers on all 152 inflations to within a second, but D2 disagrees with the actual record length on 30 of 142 records, by up to 575 s. All timings exposed by ECGBench come from the .event files, and record length from the header.
• Age and sex agree perfectly between the headers and the spreadsheet on all 520 records, so neither source needed to be preferred.

recording_type is the label; stratify_class is not. Folds are stratified on the patient’s primary occluded territory, not on the protocol phase, and that is deliberate. Because folds are patient-grouped, only patient-level attributes can actually be balanced — and every patient contributes roughly the same mix of phases, so the phase distribution comes out balanced whatever the split does, while the occluded vessel does not. The 2 left-main patients are pooled into OTHER (any class under 10 patients is), which is why stratify_class has four values and the territory table above has four rows that do not match it. Train on recording_type; primary_artery_territory keeps the unpooled value.

amplitude_range_mv is [-20, 20] here rather than ECGBench’s usual [-10, 10], and that is measured, not conventional. This is a working catheterisation laboratory: catheter and electrode movement produce transient excursions well past 10 mV that are ordinary artefact, and 71 of 520 records contain at least one sample beyond ±10 mV. At ±10 the check would discard 13.7% of the dataset for noise.

The line worth drawing is saturation, where samples were genuinely lost. The ADC rail sits at exactly ±20.48 mV (±32768 at a gain of 1600 adu/mV), and 37 records touch it. ±20 catches those plus four that came within 0.5 mV, while passing the 30 records with large but unclipped excursions.

The clipping is transient, not a dead lead: the worst-affected record (001c) has its worst lead clipped for 3.8% of samples, and no record has a lead pinned for any sustained stretch — unlike INCART, where the analogous check catches leads railed for most of a recording.

A quirk of WFDB explains why 27 records are flagged for NaN rather than amplitude: format 16 reserves −32768 as a missing-sample marker, so samples clipped at the negative rail arrive from wfdb.rdrecord as NaN, not as −20.48 mV. The two checks are catching the same physical event from opposite rails, which is why their record sets overlap in 26 of 27 cases.

089d is broken in the published release, not in transit. Its header declares 468,554 samples and its .dat holds 300,000, so wfdb.rdrecord refuses it outright. Both files match the shipped checksums, so this is upstream. The truncation falls after the balloon deflation — the inflation runs 0–278 s and the file covers 0–300 s — so the ischaemic episode itself is intact; read it with sampto=300000 if you want it. 089e has the mirror-image defect harmlessly: its header declares 300,000 samples and the .dat holds ~366,667, and wfdb simply ignores the tail.

Patients 1, 4, 5, 6 and 89 are flagged by the depositors for possible lead or sign reversal. The spreadsheet does not say which leads, so ECGBench flags all 23 of their records with suspect_leads rather than guessing. These records are not excluded from clean/ — the signals are valid, the lead identities are in doubt.