MIMIC-IV-ECG-Ext-ICD

MIMIC-IV-ECG-Ext-ICD is a label layer, not a standalone dataset: it ships no waveforms. It is one 323 MB table, records_w_diag_icd10.csv, whose 800,035 rows are exactly the 800,035 studies of MIMIC-IV-ECG, keyed by that dataset’s own study_id.

What it adds is clinical ground truth, which MIMIC-IV-ECG itself lacks: ICD-10-CM discharge diagnoses linked from the MIMIC-IV emergency-department and hospital modules. MIMIC-IV-ECG’s own labels are the ECG cart’s automated report — the machine’s opinion of the trace. These are what the patient was discharged with, which makes it possible to train an ECG model against conditions no ECG algorithm reports on.

Five diagnosis columns are shipped, differing in where the codes came from:

• ed_diag_ed — ED discharge diagnoses, from the MIMIC-IV-ED module.
• ed_diag_hosp — hospital discharge diagnoses, reached via ed_hadm_id.
• hosp_diag_hosp — hospital discharge diagnoses, reached via hosp_hadm_id.
• all_diag_hosp — the previous two, de-duplicated.
• all_diag_all — all_diag_hosp where it exists, otherwise ed_diag_ed. This is the one to train on, and the one the published benchmark uses.

Alongside them come patient demographics, date of death, the linkage ids that produced the join, and the release’s own 20-fold patient-grouped split.

Access is credentialed under the PhysioNet DUA — the same agreement as MIMIC-IV-ECG — so ECGBench never redistributes these codes.

There is deliberately no mimic_iv_ecg_ext_icd config, and no ecgbench splits --dataset mimic_iv_ecg_ext_icd. Every row here is a MIMIC-IV-ECG record, so generating a ten-fold partition for it would create a second ECGBench-blessed split over recordings that mimic_iv_ecg already partitions. A user who trained on one and evaluated on the other would be testing on training data, with both partitions carrying ECGBench’s imprimatur. Rather than create that trap and then warn about it, we do not create it.

Instead Ext-ICD is a label provider: load MIMIC-IV-ECG on ECGBench’s folds and join these columns onto it.

from ecgbench import ECGDataset
from ecgbench.labels.mimic_iv_ecg_ext_icd import load_ext_icd

ds = ECGDataset("mimic_iv_ecg", split="train", fold_numbers=[1],
                data_path="/path/to/mimic-iv-ecg/1.0/",
                metadata_source="local")
# prefix= because MIMIC-IV-ECG's own label frame also carries ecg_time.
icd = load_ext_icd("/path/to/mimic-iv-ecg-ext-icd-labels/1.0.1/", prefix="icd_")
joined = icd.reindex(ds.metadata_df["study_id"].values)   # 100% match

You need both downloads: Ext-ICD has no waveforms, and MIMIC-IV-ECG has no ICD codes. MIMIC-IV-ECG’s fold CSVs are themselves not published — see that dataset’s page for the generate-and-verify recipe — so metadata_source="local" is required here too.

The release’s own 20 folds are not ECGBench’s 10. fold and strat_fold run 0–19 and are patient-grouped, and the upstream benchmark uses folds 0–17 for training, 18 for validation and 19 for test. Use upstream_fold_split(icd, "test", prefix="icd_") to reproduce published numbers, or ECGBench’s split= to stay on ECGBench’s folds — never both. The two partitions are statistically independent, so the upstream test fold’s 39,569 records land 79.3% inside ECGBench’s train split, and 6,449 of its 8,067 patients appear there.

Recomputed from the shipped file, verified against the release’s own SHA256SUMS.txt (records_w_diag_icd10.csv → 834586ff…), so what follows is the data as published rather than a damaged local copy.

The record and patient counts are the full release, not a subset. All 800,035 study_ids and all 161,352 subject_ids of MIMIC-IV-ECG are present, with none missing and none extra in either direction. file_name equals record_list.csv’s path under a mimic-iv-ecg-diagnostic-electrocardiogram-matched-subset-1.0/ prefix in 100% of rows, and ecg_time agrees in 100% of rows. So the join is exact — which makes the coverage figure below the one that matters.

Only 468,005 of the 800,035 records carry any diagnosis — 58.5%. The other 332,030 were not part of an ED or hospital stay that MIMIC-IV holds a discharge diagnosis for. Their five diagnosis columns hold empty lists, not nulls, so an empty list is a real value here (“no linked discharge diagnosis”) and filtering on notna() will not find them. Over the labelled records all_diag_all carries a mean of 13.9 codes and a maximum of 40 — this is firmly multi-label data, and the per-column counts above do not sum to anything meaningful.

Note the 123-record gap between the 468,128 records where ecg_taken_in_ed_or_hosp is True and the 468,005 that actually carry a code: being inside a stay does not guarantee a coded discharge diagnosis.

The published 1,076-code label set reproduces exactly. The paper’s label set comes from all_diag_all by truncating each code to five characters, stripping trailing ICD-10 placeholder Xs, propagating every superclass of three characters or more, and keeping codes that appear in at least 2,000 records. Running that over the shipped table gives 1,076 codes — 361 of three characters, 466 of four, 249 of five — against the paper’s 1,076.

The X-stripping step is the one worth knowing about: W19XXXA (unspecified fall) truncates to W19XX, and only after stripping does it become the W19 category the benchmark counts. Skip it and the set comes out 1,089 codes, which looks close enough to be mistaken for a rounding difference and is not. label_set() implements the published construction; the raw 15,197 distinct codes in all_diag_all are what you get before any of it.

The label set is not a class breakdown. Because superclasses are propagated, a record coded I2510 counts toward I25, I251 and I2510 alike — which is why E78 (183,139 records) outranks I10 (181,341) after propagation while I10 leads on the raw codes. Every count here is record-level: a code appearing twice in one record counts once.

Two de-identification artefacts in the demographics. gender marks missing values with the literal string "missing" (4,489 records, the same rows whose age, anchor_age and anchor_year are null); ECGBench’s loader converts it to NaN, which is lossless because the column has no genuine nulls. And MIMIC-IV sets anchor_age to 91 for every patient older than 89, per its own documentation — which shows up here as 26,267 records at exactly 91, also the file’s maximum anchor_age. age is anchor_age plus years elapsed and so reaches 101. Do not read an age above 89 as a real age. Timestamps are shifted into the future for the same reason.

Both fold columns are leakage-free on their own terms. 0 of 161,352 subjects span more than one fold, and 0 span more than one strat_fold. fold is random and gives 39,074–40,835 records per fold; strat_fold is multi-label stratified and much more even at 39,964–40,498. v1.0.1’s release note says only that it “fixed issues with stratified folds (not used in the original benchmark)”, so fold is what reproduces published numbers.