PhysioNet/CinC Challenge 2021

The public training set of the 2021 PhysioNet/Computing in Cardiology Challenge, Will Two Do? Varying Dimensions in Electrocardiography: 88,253 twelve-lead records pooled from eight source cohorts across China, Russia, Germany and the USA, redistributed in one uniform WFDB tree.

This is a meta-dataset, and that is the single most important thing to know about it. Five of the eight cohorts are datasets ECGBench catalogues separately, and the records were renamed, so the overlap is invisible from the identifiers. Four of those overlaps are verified against the files in the relationships above — including PTB-XL, which is present in full and bit-identical. Training here and evaluating on PTB-XL, PTBDB, INCART, CPSC-2018, Chapman-Shaoxing, Ningbo or PhysioNet ecg-arrhythmia is testing on training data.

Only the training set was ever published. The challenge’s hidden validation and test sets are not in this release, which is why the record count is 88,253 and not the ~131,000 the challenge described. There is consequently no reproducible official split, and ECGBench generates its own 10 folds.

Sampling rate and record length are per-record properties here, unlike every other dataset in this catalogue. Rate is 500 Hz for 87,663 records, 1000 Hz for the 516 ptb ones and 257 Hz for the 74 st_petersburg_incart ones; length runs from 5 s to 1800 s, with 1,650 distinct lengths in the cpsc_2018 cohort alone. Both are exposed as label columns. Records cannot be batched as they are — take a fixed window=(start, length), or use batch_size=1.

Labels are multi-label SNOMED-CT codes from the #Dx header field: 133 distinct codes, 2.06 per record on average and up to 12, with no unlabelled record. No code table ships with the data, so ECGBench packages the challenge’s own.

Records: 88,253, not ~131,000. The challenge description quoted the full corpus including the hidden validation and test sets; PhysioNet only ever published the training set. Recomputed as the number of .hea files under training/. An earlier version of this catalogue page said “~130,862” — that figure belongs to the challenge, not to this download.

Patients: not published. No patient identifier ships with any cohort and records were renamed, so patient counts are unrecoverable from these files. This is exact for six cohorts (one record per patient), but not for ptb and st_petersburg_incart, whose source datasets have several recordings per patient — 113 of PTBDB’s 290 patients and roughly 32 patients across INCART’s 75 records. Those 590 records (0.67%) can therefore place one patient in more than one fold. Every other dataset in ECGBench with repeated patients groups folds by patient; this one cannot.

The diagnosis table is multi-label and does not sum to the record total. It counts records carrying each code, over 181,817 code instances across 88,253 records. Derived from the #Dx field of all 88,253 headers, joined to the challenge’s code table. The Scored column marks the 30 classes the challenge metric evaluated (as 26 classes, four pairs being scored as equivalent); 81,966 records carry at least one scored code and 6,287 carry none. Every one of the 133 codes present in the data has a table entry, and the per-code totals agree with the official dx_mapping_*.csv Total column on all 133 codes.

The stratification label is not the diagnosis. Because #Dx ordering carries no clinical meaning in this release — PTB-XL’s lists are numerically sorted, Chapman’s are not — there is no primary diagnosis to read off. ECGBench derives stratify_dx as the globally rarest code each record carries, ties broken on the lowest numeric code, purely so stratified folds are well defined. It keeps all 133 classes representable, where taking the first listed code collapses to 123. Train on dx, never on stratify_dx.

The code table is packaged with ECGBench, not with the data. No mapping ships in the download, so ecgbench/data/challenge2021_dx_mapping.csv holds the concatenation of the challenge’s own dx_mapping_scored.csv (30) and dx_mapping_unscored.csv (103) from the official scoring code (evaluation-2021, BSD-2-Clause). It covers exactly the 133 codes present, with no duplicate code or abbreviation.

One record is misnamed. The ningbo cohort contains S23074 where every other record in that cohort is JS##### (it is JS23074 in the ecg-arrhythmia release). Record names are still globally unique, so nothing collides, but a join on names against ecg-arrhythmia silently drops it.

Two genuine defects, both concentrated in specific cohorts rather than spread evenly — another reason not to treat the eight sources as interchangeable.

All-zero leads (1,531 records). Entire leads recorded as exactly 0.0, most often the precordials: 976 records have all six of V1–V6 flat. Overwhelmingly ningbo (1,480 of the 1,531). Verified on disk rather than inferred — A0718’s lead 11 is all zeros in the file.

Amplitude outliers (1,021 records). Excursions past ±10 mV. Many sit exactly at ±32.767 mV, the 16-bit rail, so they are saturation rather than physiology. st_petersburg_incart is worst hit at 28 of its 74 records (37.8%). 105 records also contain NaN samples, and 892 records are excluded for an amplitude outlier alone.

Per-cohort exclusion rates: st_petersburg_incart 37.8%, ningbo 6.3%, cpsc_2018_extra 1.8%, cpsc_2018 1.7%, chapman_shaoxing 0.4%, ptb 0.4%, georgia 0.3%, ptb-xl 0.2%. clean/ drops them; original/ keeps them with the reason in quality_issues.