Symile-MIMIC

Symile-MIMIC is a multimodal cohort over MIMIC-IV-ECG’s recordings, not a standalone ECG dataset. It pairs each of 11,622 hospital admissions with three modalities drawn from MIMIC-IV:

• a 12-lead ECG taken within 24 h of admission, from MIMIC-IV-ECG — 10 s at 500 Hz, 5,000 samples per lead;
• a chest X-ray taken in the 24–72 h window post-admission, from MIMIC-CXR-JPG, with the 14 CheXpert findings for it;
• up to 50 blood labs drawn within 24 h of admission, from MIMIC-IV hosp.

It was built for the NeurIPS 2024 Symile paper, whose contrastive objective needs three or more modalities at once, and its evaluation task is CXR retrieval: given a query’s ECG and labs, pick that query’s chest X-ray out of 10 candidates. Only admissions with all three modalities present — and a PA or AP view, and at least one of the 50 labs — qualified, which is why 11,622 admissions come out of MIMIC-IV-ECG’s 800,035 studies.

Access is credentialed under the PhysioNet Credentialed Health Data Use Agreement — the same agreement as MIMIC-IV-ECG — so ECGBench redistributes none of it.

There is deliberately no symile_mimic config, and no ecgbench splits --dataset symile_mimic. Every ECG here is a MIMIC-IV-ECG record, so generating a ten-fold partition 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 Symile-MIMIC is a cohort and label provider: load MIMIC-IV-ECG on ECGBench’s folds and join these columns onto it.

from ecgbench import ECGDataset
from ecgbench.labels.symile_mimic import by_study_id, load_cohort

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 subject_id
# and ecg_time.
cohort = load_cohort("/path/to/symile-mimic/1.0.0/", prefix="sym_")
joined = by_study_id(cohort, prefix="sym_").reindex(
    ds.metadata_df["study_id"].values)

You need both downloads: Symile-MIMIC has no ECGBench folds, and MIMIC-IV-ECG has none of these columns. 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 join is deliberately partial. Symile-MIMIC covers 11,610 of MIMIC-IV-ECG’s 800,035 studies, so a single ECGBench fold of 78,655 records matches 1,135 of them (1.44%). That is the cohort being a cohort, not a broken join — unlike MIMIC-IV-ECG-Ext-ICD, which covers all 800,035 records and should match 100%.

The release’s own train/val/test split is not ECGBench’s. Symile’s splits are patient-disjoint on their own terms — 0 subjects are shared between any pair — but they are statistically independent of ECGBench’s folds: 75.6% of the 464 test studies land in ECGBench’s train split, and 349 of the 461 test subjects appear there. load_split() gives you the release’s partition for reproducing published numbers; ECGDataset(split=...) gives you ECGBench’s. Never mix them.

Recomputed from the shipped files. All 40 files verify against the release’s own SHA256SUMS.txt, so what follows is the data as published rather than a damaged or filtered local copy.

11,622 is admissions; the distinct ECG count is 11,610. The row unit here is the hospital admission (hadm_id, unique across all 11,622 rows), not the recording. 12 ECG studies each serve two admissions — always two admissions of the same patient hours or a day apart, where the same ECG fell within 24 h of both. The records figure above is the release’s own headline count; if you are counting recordings, the number is 11,610. The CXR side is ragged the same way, at 11,609 distinct DICOMs. And 1,476 of the 9,573 patients contribute more than one admission, up to 13.

The column literally named study_id is the CXR’s, not the ECG’s. It is byte-identical to cxr_study_id in 100% of rows, while MIMIC-IV-ECG’s key lives in ecg_study_id. Joining MIMIC-IV-ECG on the shipped study_id matches nothing at all, which looks like a broken dataset rather than a wrong column name. ECGBench’s loader drops the duplicate so the trap is gone. The split CSVs drop ecg_study_id as well, so it is recovered there from the last segment of ecg_path — verified equal to ecg_study_id in 100% of rows, with ecg_path itself equal to record_list.csv’s path in 100%.

CheXpert labels have four states, and two of them are not “no”. 1.0 positive, 0.0 negative, −1.0 uncertain, and NaN “the report does not mention this finding” — where NaN is the majority state for most findings. Reading −1.0 as negative, or NaN as negative, are modelling choices with different results, so ECGBench’s chexpert_targets() asks you to pick rather than defaulting silently.

This is multi-label data, so the columns do not sum to 11,622. No Finding is never 0.0 or −1.0, making it 1,606 positives against 10,016 unmentioned.

The split CSVs are not a partition of the full table, and they drop columns. 10,000 + 750 + 464 = 11,214 of the 11,622 admissions appear in a split; the other 408 were discarded by the release’s patient-disjointness filter, which drops any val or test candidate whose patient already appears in an earlier split. The split CSVs also carry only 6 of the 14 CheXpert labels — Atelectasis, Cardiomegaly, Edema, Lung Opacity, No Finding, Pleural Effusion — and no demographics, ecg_study_id, ecg_time or CXR metadata. Those exist only in symile_mimic_data.csv.

val.csv’s label and label_hadm_id columns carry no information. They are constant (label == 1, label_hadm_id == hadm_id for all 750 rows), a side effect of the release’s create_dataset_splits.py mutating the validation frame in place while building val_retrieval.csv from it. val is a plain validation set, and its data_npy/ directory correctly ships no label tensor.

Lab coverage is uneven, but there are no sentinel rails. 16 of the 50 labs are present for ≥99% of admissions, 18 for 50–90%, and 16 for under half; the three itemids MIMIC-IV names H, L and I (50934, 51678, 50947) bring up the rear at 10.5%. Every admission has at least one lab, which is what the release’s labs_all_nan column records — it is 0 for all 11,622 rows. Unlike MIMIC-IV-ECG’s machine_measurements.csv, which encodes “unmeasurable” as integer rails, missing labs here are genuine NaNs, so notna() is the right test. A handful of values are nonetheless physiologically impossible rather than missing — 10 of 5,254 Base Excess values fall below −30, the minimum being −413 — and are MIMIC-IV data-entry errors passed through unchanged.

The data_npy ECG tensors are not millivolts, and cannot be converted back. Each record was min-max normalised to [−1, 1] over all 12 leads at once — 2 * (x − x.min()) / (x.max() − x.min()) − 1, verified to float32 precision against wfdb.rdrecord on the corresponding MIMIC-IV-ECG record — and the per-record min and max are not shipped, so the transform is not invertible. They are also stored channel-last with a leading singleton, (n, 1, 5000, 12), rather than ECGBench’s (12, 5000). For real millivolts, read MIMIC-IV-ECG through ECGDataset. Note also that the release’s README describes these files as .pt, because its script saves torch.save output; PhysioNet ships .npy.

Lead order is MIMIC-IV-ECG’s, so aVF comes before aVL. signal[4] is aVF here, not aVL as in PTB-XL or Chapman. Pass leads= to ECGDataset to select by name.

Timestamps and ages are de-identification artefacts. MIMIC-IV shifts dates into the future — ecg_time spans 2110 to 2208 — and caps anchor_age at 91 for every patient over 89, which shows up as 580 rows at exactly 91, the file maximum. age is anchor_age plus years elapsed and so reaches 100 (median 69). Do not read an age above 89 as a real age.

47 of the 11,610 ECG studies are excluded from ECGBench’s clean/ version. The release’s own pipeline already removed ECGs with NaN values or an all-zero signal, but ECGBench’s validation additionally flags 36 amplitude_outlier records — leads railing near ±20 mV — and 11 with a missing_lead, with no record failing both. They remain in original/, flagged.