MIMIC-IV-ECG - Diagnostic Electrocardiogram Matched Subset

Abstract

The MIMIC-IV-ECG module contains approximately 800,000 diagnostic electrocardiograms across nearly 160,000 unique patients. These diagnostic ECGs use 12 leads and are 10 seconds in length. They are sampled at 500 Hz. This subset contains all of the ECGs for patients who appear in the MIMIC-IV Clinical Database. When a cardiologist report is available for a given ECG, it is also provided. The patients in MIMIC-IV-ECG have been matched against the MIMIC-IV Clinical Database, making it possible to link to information across the MIMIC-IV modules. 

Background

An Electrocardiogram or ECG / EKG measures the electrical activity associated with the heart [1]. Diagnostic ECGs are a standard part of a patients care [2]. The standard ECG leads are denoted as lead I, II, III, aVF, aVR, aVL, V1, V2, V3, V4, V5, V6. They are routinely obtained when admitted to the Emergency Department or to a hospital floor. ECGs will typically be repeated for patients who exhibit cardiac symptoms such as chest pain or abnormal rhythms. Daily ECGs may be obtained following acute cardiovascular events such as myocardial infarction. Patients in the Intensive Care Unit (ICU) are continuously monitored to detect rhythm abnormalities, but full ECGs are needed to evaluate evidence of cardiac ischemia or infarction. However, diagnostic ECGs typically only comprise a small part of understanding the overall condition of a subject at the hospital. To fully understand how to best treat a given patient, a broader set of data is collected which may include: patient demographics, diagnosis, medications, lab tests, and additional information. This broader set of clinical information is shared as part of the MIMIC-IV Clinical Database [3]. The MIMIC-IV-ECG Matched Subset contains the vast majority of diagnostic ECGs collected between 2008 - 2019 for subjects in MIMIC-IV.

Methods

As part of routine care, diagnostic ECGs are collected across Beth Israel Deaconess Medical Center (BIDMC). Three types of information associated with an ECG are presented here. The electrocardiogram waveforms themselves, the machine measurements (ex: average RR interval as calculated by the machine), and the cardiologist reports. Identifiers connected to the ECGs allow this information to be connected back to the patients overall electronic health record. All of the information is de-identified to satisfy the US Health Insurance Portability and Accountability Act of 1996 (HIPAA) Safe Harbor requirements.

Electronic Health Record

Patients from the MIMIC-IV Clinical Database who had ECGs collected between 2008 - 2019 are included as part of MIMIC-IV-ECG. The diagnostic ECGs are collected on machines from various manufacturers. When the ECG is collected, the machine is populated with the patient's demographics and their medical record number (MRN).

As part of de-identification the raw identifiers are shifted. The patient's MRN was used to match a given 12-lead ECG record to the corresponding subject ID in the MIMIC-IV Clinical Database. As another part of the de-identification, the date-time information was shifted to obscure the actual date and time. Relative date-time information for a given subject is preserved though. The shifted date-times were matched against date-times in the subject's MIMIC-IV Clinical Database records. A unique study_id was generated for each record.

Electrocardiogram Waveforms

If a patient appears in the MIMIC-IV Clinical Database, all of their available ECG waveforms were pulled. This includes ECGs from the BIDMC emergency department, hospital (including the ICU), and outpatient care centers. We converted the ECGs from the manufacturers format to the open WFDB format [4] with each WFDB record comprised of a header (.hea) file and a signal (.dat) file. The files were then transferred from BIDMC to MIT for additional processing.

We scrubbed the WFDB header files for PHI such that only the signal information, subject ID, and shifted date-time are provided. Timestamps for events in the MIMIC-IV Clinical Database, such as drug administration, are aligned with the timestamps in MIMIC-IV-ECG. However, some of the diagnostic ECGs provided here were collected outside of ED or ICU visits at the hospital. Since the MIMIC-IV Clinical Database is comprised solely of ED and ICU data, the ECG timestamp can occur before or after a visit from the clinical database.

Machine Measurements

The ECG machine generates summary reports and summary measures (ex: RR interval, QRS onset and end, etc.) for each diagnostic ECG. We collectively refer to these as machine measurements. The machine output is parsed and any PHI is removed. In particular, the MRN is shifted to subject_id, the de-identified study_id is assigned in a manner consistent with the ECG waveform files, and the raw Cart ID is randomly shifted to create a de-identified cart_id. There was no PHI in the report lines. 

The global machine measures are provided in this release. These global measures are calculated across all 12 leads. Machine measurements for individual leads may be released in a future version of this project. 

Cardiologist Reports

Most ECG waveforms get read by a cardiologist and an associated report is generated from the reading. These reports are provided where available.

These ECG reports were de-identified using a rule-based approach [5, 6, 7], similar to that used for other MIMIC reports. Each instance of PHI was replaced by three underscores.

Data Description

Electrocardiogram Waveforms

Approximately 800,000 ten-second-long 12 lead diagnostic ECGs across nearly 160,000 unique subjects are provided in the MIMIC-IV-ECG module. Around 5% of the available diagnostic ECGs were withheld from this release so they can be used as a hidden test set in workshops and challenges. The ECGs are sampled at 500 Hz. The patients in this module have been matched with the MIMIC-IV Clinical Database. Many of the provided diagnostic ECGs overlap with a MIMIC-IV hospital or emergency department stay but a number of them do not overlap.  All available diagnostic ECGs for a particular patient have been placed under a single subdirectory (pXXXXXXXX), named according to the patient's MIMIC-IV subject ID, provided as XXXXXXXX. These subdirectories are further divided into intermediate group level directories based on the range of subject IDs contained within. Each group directory contains subject IDs within a range of 1000 potential ID values, pNNNN. For example, the p1000 group level directory contains all subject IDs between 1000000 and 1000999, while the p1025 group level directory contains all subject ID's between 1025000 and 1025999. 

Each waveform record path is named as files/pNNNN/pXXXXXXXX/sZZZZZZZZZ/ZZZZZZZZZ, where NNNN is the group level directory, XXXXXXXX is the subject ID, and ZZZZZZZZZ is the study ID. An example of the file structure is as follows:

files
├── p1000
|   └── p10001725
|       └── s102147240
|           ├── 102147240.dat
|           └── 102147240.hea
└── p1002
    └── p10023771
        ├── s104496507
        │   ├── 104496507.dat
        │   └── 104496507.hea
        ├── s108135749
        │   ├── 108135749.dat
        │   └── 108135749.hea
        └── s105384473
            ├── 105384473.dat
            └── 105384473.hea

Above we find two subjects p10001725 (under the p1000 group level directory) and p10023771 (under the p1002 group level directory). For subject p10001725 we find one study: s102147240. For p10023771 we find three studies: s104496507, s108135749, s105384473. The study identifiers are completely random, and their order has no implications for the chronological order of the actual studies. Each study has a like named .hea and .dat file, comprising the WFDB record. 

The record_list.csv file contains the file name and path for each WFDB record. It also provides the corresponding subject ID and study ID. The subject ID can be used to link a subject from MIMIC-IV-ECG to the other modules in the MIMIC-IV Clinical Database. 

Machine Measurements

Machine measurements for each ECG waveform are provided in the machine_measurements.csv file. A data dictionary provides a description for each of the columns in machine_measurements_data_dictionary.csv. The machine measurements table provides the machine generated reports in columns report_0..report_17. The report lines are provided as generated by the machine. In some cases there will be a column with no text in between columns with text (ex: report_0: <text_a>, report_1: empty, report_2: <text_b>). In addition to the summary measurements (rr_interval, qrs_onset, qrs_end, etc.) columns for the machine's bandwidth and filter settings (filtering) are provided. A cart_id is provided which can be used to track which machine was used for a given ECG. Finally, the subject_id, study_id, and date are provided, consistent with the ECG waveform files themselves. 

Cardiologist Reports 

A little more than 600,000 cardiologist reports are available for the ~800,000 diagnostic ECGs. Not all diagnostic ECGs get read by a cardiologist. This is the primary reason that there are fewer reports than waveforms.

The provided reports.csv table has a text column which contains the de-identified cardiologist report for a given diagnostic ECG. This table also contains the subject ID, study ID, and waveform path. This information can be used to connect a report to a given subject and their diagnostic ECG waveform. Each report gets a unique ID, which is composed of the subject ID, the abbreviation for the domain (EK) that the report comes from, and a sequential integer. The sequential integer is also listed in its own column and can be used to decipher the order in which ECGs were collected for a given subject.

The information from the reports.csv table is also available on BigQuery [8].

Usage Notes

This module provides MIMIC-IV users an additional, potentially important piece of information for their research using MIMIC. 

A limitation of this dataset is that the 12-lead ECG timestamps may not be perfectly time synced with the other waveforms in MIMIC, as they are collected from different machines. An additional limitation, as noted above, is that some of the ECGs provided here were collected outside of the ED and ICU at the hospital. This means that the timestamps for those ECGs won't overlap with data from the MIMIC-IV Clinical Database.

The signals can be viewed in Lightwave by clicking the Visualize waveforms links in the Files section below. Additionally, the signals can be read by using the WFDB toolboxes provided on PhysioNet: WFDB (in C) [9], WFDB-Matlab [10], and WFDB-Python [11]. Here is a basic script for reading a downloaded record from this project and plotting it by using the WFDB-Python toolbox:

import wfdb 
rec_path = '/files/p1000/p10001725/s102147240/102147240' 
rd_record = wfdb.rdrecord(rec_path) 
wfdb.plot_wfdb(record=rd_record, figsize=(24,18), title='Study 102147240 example', ecg_grids='all')

where rec_path is the path to the name of the .hea and .dat files for the record you'd like to plot.

Here we provide an example of how subject p10023771 from MIMIC-IV-ECG can be linked to their admission information in the MIMIC-IV Clinical Database.  Executing this from BigQuery:

SELECT * FROM `physionet-data.mimiciv_hosp.admissions` WHERE subject_id=10023771

we see that the patient only has one admission to the hospital with an admittime = 2113-08-25T07:15:00 and a dischtime = 2113-08-30T14:15:00. We also need to check to see if they were seen in the emergency department and not admitted to the hospital:

SELECT * FROM `physionet-data.mimiciv_ed.edstays` WHERE subject_id = 10023771

We observe that they did not have a stay in the emergency department.

Next, we get the timestamps from the diagnostic ECGs by checking the base_date and base_time variables. These are the variables used in the WFDB format for storing date and time. They correspond with the timestamps for the diagnostic ECGs that are provided in the summary tables. We then save the result to a csv file:

from pathlib import Path
import pandas as pd

import wfdb

# get the path to all the study .hea files for p10023771
paths = list(Path("p10023771/.").rglob("*.hea"))

# get date and time for each study
date_times = {'study':[],'date':[],'time':[]} # use a dictionary to store the date and time for each study
for file in paths:
    study = file.stem
    metadata = wfdb.rdheader(f'{file.parent}/{file.stem}')
    date_times['study'].append(study)
    date_times['date'].append(metadata.base_date)
    date_times['time'].append(metadata.base_time)

df_date_times = pd.DataFrame(data=date_times)
df_date_times.to_csv('p10023771_date_times.csv', index=False)

We observe the following for the 3 diagnostic ECGs for p10023771: 

where the date is given before the T as YYYY-MM-DD and the time is given after the T as HH:MM. Comparing this to the subjects admission in the MIMIC-IV Clinical Database:

we observe that s104496507 and s108135749 occurred prior to their only hospital admission while s105384473 occurred during their hospital admission. 

We can also check the available cardiologist reports for this subject by running this command in BigQuery:

SELECT * FROM `lcp-consortium.mimic_ecg.reports` WHERE subject_id = 10023771

We find that there are cardiologist reports available for s108135749 and s105384473 but not s104496507.

Release Notes

MIMIC-IV-ECG v0.3

This release provides a table for the diagnostic ECG machine measurements. 

Ethics

The project was approved by the Institutional Review Boards of Beth Israel Deaconess Medical Center (Boston, MA) and the Massachusetts Institute of Technology (Cambridge, MA). Requirement for individual patient consent was waived because the project did not impact clinical care and all protected health information was deidentified.

Acknowledgements

SH, RM, BG, and TP are funded by the Massachusetts Life Sciences Center, Nov. 30, 2020. NG is supported by National Institutes of Health National Library of Medicine Biomedical Informatics and Data Science Research Training Program under grant number T15LM007092-30. BG, TP, AJ, BM, CF, DM, and RM are supported by the National Institute of Biomedical Imaging and Bioengineering (NIBIB) under NIH grant number R01EB030362.

Conflicts of Interest

The author(s) have no conflicts of interest to declare.

References

1. Geselowitz DB. On the theory of the electrocardiogram. Proceedings of the IEEE. 1989 Jun;77(6):857-76.
2. Harris PR. The Normal electrocardiogram: resting 12-Lead and electrocardiogram monitoring in the hospital. Critical Care Nursing Clinics. 2016 Sep 1;28(3):281-96.
3. Johnson, A., Bulgarelli, L., Pollard, T., Horng, S., Celi, L. A., & Mark, R. (2021). MIMIC-IV (version 1.0). PhysioNet. https://doi.org/10.13026/s6n6-xd98.
4. Documentation for the Waveform Database (WFDB) file format. https://wfdb.io/ [Accessed 21 June 2022]
5. Margaret Douglass, Computer-assisted de-identification of free-text nursing notes. Master's Thesis, 2005. MIT.
6. Neamatullah, I., Douglass, M.M., Lehman, L.H., Reisner, A., Villarroel, M., Long, W.J., Szolovits, P., Moody, G.B., Mark, R.G., Clifford, G.D. (2007). De-Identification Software Package (version 1.1). PhysioNet. doi:10.13026/C20M3F
7. Neamatullah I, Douglass MM, Lehman LH, Reisner A, Villarroel M, Long WJ, Szolovits P, Moody GB, Mark RG, Clifford GD. Automated de-identification of free-text medical records. BMC medical informatics and decision making. 2008 Dec;8(1):1-7. doi:10.1186/1472-6947-8-32
8. Documentation about using the Medical Information Mart for Intensive Care (MIMIC) Database with Google BigQuery. https://mimic.mit.edu/docs/gettingstarted/cloud/ [Accessed 21 June 2022]
9. Documentation for the Waveform Database (WFDB) toolbox in C. https://physionet.org/content/wfdb/10.7.0/ [Accessed 21 June 2022]
10. Documentation for the Waveform Database (WFDB) toolbox for Matlab. https://physionet.org/content/wfdb-matlab/0.10.0/ [Accessed 21 June 2022]
11. Documentation for the Waveform Database (WFDB) toolbox for Python. https://physionet.org/content/wfdb-python/3.4.1/ [Accessed 21 June 2022]