PhysioNet/CinC Challenge 2020

The public training set of the 2020 PhysioNet/Computing in Cardiology Challenge, Classification of 12-lead ECGs: 43,101 twelve-lead records pooled from six source cohorts across China, Russia, Germany and the USA, redistributed in one uniform WFDB tree.

This release is contained entirely in Challenge 2021. Not “overlaps with” — contained. All 43,101 record names appear there, and all 43,101 waveform files are bit-identical, checked by comparing the two releases’ own published SHA256SUMS.txt. The 2021 challenge simply added the Chapman-Shaoxing and Ningbo cohorts on top (88,253 − 45,152 = 43,101). Only 705 of the 43,101 headers differ at all, and those differences are label housekeeping described below. Never train on one year and evaluate on the other.

It is also a meta-dataset, which is the second thing to know. Four of the six cohorts are datasets ECGBench catalogues separately, and the records were renamed, so the overlap is invisible from the identifiers. Training here and evaluating on PTB-XL, PTBDB, INCART or CPSC-2018 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 43,101 and not the ~52,500 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. Rate is 500 Hz for 42,511 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: 111 distinct codes, 2.18 per record on average and up to 10, with no unlabelled record. No code table ships with the data, so ECGBench packages the challenge’s own.

Records: 43,101, not ~52,500. 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/. All 86,301 shipped data files were verified against the release’s own SHA256SUMS.txt before any figure on this page was computed.

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 four 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 (1.37%) 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.

631 records repeat a code inside their own #Dx list, and ECGBench deduplicates them. 596 records list 284470004 (PAC) twice, 30 list 17338001 (VPB) twice, five others repeat one code and one lists a code three times — 628 in georgia and 3 in cpsc_2018_extra. This is not cosmetic: counting raw list entries inflates Georgia’s PAC total from 639 records to 1,236 and makes the shipped v1.0.2 data look as though it disagrees with the official code table. After deduplication all 111 codes and all six per-cohort columns of the official dx_mapping_*.csv reproduce exactly (93,843 code-record pairs). The Challenge 2021 re-release of the same records had already deduplicated and numerically sorted these lists, which is why the 2021 page does not mention this.

The diagnosis table is multi-label and does not sum to the record total. It counts records carrying each code, over 93,843 code instances across 43,101 records. Derived from the #Dx field of all 43,101 headers, joined to the challenge’s code table. The Scored column marks the 27 classes the 2020 challenge metric evaluated (as 24 classes, three pairs — CRBBB/RBBB, PAC/SVPB, PVC/VPB — being scored as equivalent); 37,749 records carry at least one scored code and 5,352 carry none.

The scored subset is not Challenge 2021’s. 2021 scored 30 classes: these 27 plus PRWP (365413008) and CLBBB (733534002), neither of which occurs anywhere in this release, and BBB (6374002), which does occur here (137 records) but was unscored in 2020. Nothing scored in 2020 was dropped in 2021. Use the table packaged for the year you are reporting against.

The stratification label is not the diagnosis. Because #Dx ordering carries no clinical meaning in this release — it varies by cohort and, in Georgia, is not internally consistent — 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 111 classes representable, where taking the first listed code collapses to 102. 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/challenge2020_dx_mapping.csv holds the concatenation of the challenge’s own dx_mapping_scored.csv (27) and dx_mapping_unscored.csv (84) from the official scoring code (evaluation-2020, BSD-2-Clause). It covers exactly the 111 codes present, with no duplicate code or abbreviation.

Age carries two sentinel values. 181 records have no age at all; beyond that, 204 ptb-xl records record 300 (PTB-XL’s own convention for a patient older than 89) and 6 CPSC records record -1. Genuine ages run 1–92. ECGBench leaves all three states in the age column rather than collapsing them, and exposes the sentinels as ecgbench.labels.challenge2020.AGE_SENTINELS. Sex is normalised: the 74 st_petersburg_incart records spell it M/F where the other five cohorts spell it Male/Female, and the loader maps them onto the long form. One record has no sex.

Worth stating precisely, because it is the only difference between two datasets that hold the same recordings.

All 43,101 .mat waveform files are bit-identical across the two releases — compared via both releases’ published SHA256SUMS.txt, 43,101 of 43,101 hashes equal. Of the 43,101 headers, 42,396 are identical and 705 differ, and every one of those 705 differences is in the comment block, not the signal specification:

• 631 records — the #Dx list. 2021 deduplicated the repeated codes described above and sorted the codes numerically; 2020 ships them as written. Once ECGBench deduplicates, the sets of codes agree on all 43,101 records.
• 74 records — #Sex in the st_petersburg_incart cohort, M/F in 2020 and Male/Female in 2021. ECGBench normalises 2020 to match.

Age, sampling rate, sample count, lead count and cohort assignment are identical on all 43,101 records. So the two datasets differ in their label encoding and in the 45,152 extra records 2021 adds — not in any recording.

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

Amplitude outliers (250 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%). Eight of these records also contain NaN samples.

All-zero leads (36 records). Entire leads recorded as exactly 0.0, 85 leads in total; one record has all twelve flat.

Per-cohort exclusion rates: st_petersburg_incart 37.8%, cpsc_2018_extra 1.8%, cpsc_2018 1.7%, ptb 0.4%, georgia 0.3%, ptb-xl 0.2%. These match Challenge 2021’s rates for the same six cohorts exactly, which they must — they are the same files. clean/ drops them; original/ keeps them with the reason in quality_issues.