MIMIC-IV

Abstract

Retrospectively collected medical data has the opportunity to improve patient care through knowledge discovery and algorithm development. Broad reuse of medical data is desirable for the greatest public good, but data sharing must be done in a manner which protects patient privacy. Here we present Medical Information Mart for Intensive Care (MIMIC)-IV, a large deidentified dataset of patients admitted to the emergency department or an intensive care unit at the Beth Israel Deaconess Medical Center in Boston, MA. MIMIC-IV contains data for over 65,000 patients admitted to an ICU and over 200,000 patients admitted to the emergency department. MIMIC-IV incorporates contemporary data and adopts a modular approach to data organization, highlighting data provenance and facilitating both individual and combined use of disparate data sources. MIMIC-IV is intended to carry on the success of MIMIC-III and support a broad set of applications within healthcare.

Background

In recent years there has been a concerted move towards the adoption of digital health record systems in hospitals. In the US, nearly 96% of hospitals had a digital electronic health record system (EHR) in 2015 [1]. Retrospectively collected medical data has increasingly been used for epidemiology and predictive modeling. The latter is in part due to the effectiveness of modeling approaches on large datasets [2]. Despite these advances, access to medical data to improve patient care remains a significant challenge. While the reasons for limited sharing of medical data are multifaceted, concerns around patient privacy are highlighted as one of the most significant issues. Although patient studies have shown almost uniform agreement that deidentified medical data should be used to improve medical practice, domain experts continue to debate the optimal mechanisms of doing so. Uniquely, the MIMIC-III database adopted a permissive access scheme which allowed for broad reuse of the data [3]. This mechanism has been successful in the wide use of MIMIC-III in a variety of studies ranging from assessment of treatment efficacy in well defined cohorts to prediction of key patient outcomes such as mortality. MIMIC-IV aims to carry on the success of MIMIC-III, with a number of changes to improve usability of the data and enable more research applications.

Methods

MIMIC-IV is sourced from two in-hospital database systems: a custom hospital wide EHR and an ICU specific clinical information system. The creation of MIMIC-IV was carried out in three steps:

1. Acquisition. Data for patients who were admitted to the BIDMC emergency department or one of the intensive care units were extracted from the respective hospital databases. A master patient list was created which contained all medical record numbers corresponding to patients admitted to an ICU or the emergency department between 2008 - 2022. All source tables were filtered to only rows related to patients in the master patient list.
2. Preparation. The data were reorganized to better facilitate retrospective data analysis. This included the denormalization of tables, removal of audit trails, and reorganization into fewer tables. The aim of this process is to simplify retrospective analysis of the database. Importantly, data cleaning steps were not performed, to ensure the data reflects a real-world clinical dataset.
3. Deidentify. Patient identifiers as stipulated by HIPAA were removed. Patient identifiers were replaced using a random cipher, resulting in deidentified integer identifiers for patients, hospitalizations, and ICU stays. Structured data were filtered using look up tables and allow lists. If necessary, a free-text deidentification algorithm was applied to remove PHI from free-text. Finally, date and times were shifted randomly into the future using an offset measured in days. A single date shift was assigned to each subject_id. As a result, the data for a single patient are internally consistent. For example, if the time between two measures in the database was 4 hours in the raw data, then the calculated time difference in MIMIC-IV will also be 4 hours. Conversely, distinct patients are not temporally comparable. That is, two patients admitted in 2130 were not necessarily admitted in the same year.

After these three steps were carried out, the database was exported to a character based comma delimited format.

Data Description

MIMIC-IV is grouped into two modules: hosp, and icu. Organization of the data into these modules reflects their provenance: data in the hosp module is sourced from the hospital wide EHR, while data in the icu module is sourced from the in-ICU clinical information system (MetaVision). A total of 364,627 unique individuals are in MIMIC-IV v3.0, each represented by a unique subject_id. These individuals had 546,028 hospitalizations and 94,458 unique ICU stays.

hosp

The hosp module contains detailed data regarding 546,028 unique hospitalizations for 223,452 unique individuals. Measurements in the hosp module are predominantly recorded during the hospital stay, though some tables include data from outside an admitted hospital stay as well (e.g. outpatient or emergency department laboratory tests in labevents). Patient demographics (patients), hospitalizations (admissions), and intra-hospital transfers (transfers) are recorded in the hosp module. Other information in the hosp module includes laboratory measurements (labevents, d_labitems), microbiology cultures (microbiologyevents, d_micro), provider orders (poe, poe_detail), medication administration (emar, emar_detail), medication prescription (prescriptions, pharmacy), hospital billing information (diagnoses_icd, d_icd_diagnoses, procedures_icd, d_icd_procedures, hcpcsevents, d_hcpcs, drgcodes), online medical record data (omr), and service related information (services).

Provider information is available in the provider table. The provider_id column is a deidentified character string which uniquely represents a single care provider. As provider_id is used in different contexts across the module, a prefix is usually present in data tables to contextualize how the provider relates to the event. For example, the provider who admits the patient to the hospital is documented in the admissions table as subject_id. All columns which have a suffix of provider_id may be linked to the provider table.

Deidentified dates and aligning stays to year groups

All dates in MIMIC-IV have been deidentified by shifting the dates into a future time period between 2100 - 2200. This shift is done independently for each patient, and as a result two patients admitted in the deidentified year 2120 cannot be assumed to be admitted in the same year. To provide information about the original time period when a patient was admitted, the patients table provides a set of columns with the "anchor_" prefix. The anchor_year column is a deidentified year occurring sometime between 2100 - 2200, and the anchor_year_group column is one of the following values: "2008 - 2010", "2011 - 2013", "2014 - 2016", "2017 - 2019", and "2020 - 2022". These pieces of information allow researchers to infer the approximate year a patient received care. For example, if a patient's anchor_year is 2158, and their anchor_year_group is 2011 - 2013, then any hospitalizations for the patient occurring in the year 2158 actually occurred sometime between 2011 - 2013. In order to minimize accidental release of information, only a single anchor_year is provided per subject_id. Consequently, individual stays must be aligned to the anchor year using the respective date (e.g. admittime). Finally, the anchor_age provides the patient age in the given anchor_year. If the patient was over 89 in the anchor_year, this anchor_age has been set to 91 (i.e. all patients over 89 have been grouped together into a single group with value 91, regardless of what their real age was).

Out of hospital linkage of date of death

Date of death is available within the dod column of the patients table. Date of death is derived from hospital records and state records. If both exist, hospital records take precedence. State records were matched using a custom rule based linkage algorithm based on name, date of birth, and social security number. State and hospital records for date of death were collected two years after the last patient discharge in MIMIC-IV, which should limit the impact of reporting delays in date of death.

Dates of death occurring more than one year after hospital discharge are censored as a part of the deidentification process. As a result, the maximum time of follow up for each patient is exactly one year after their last hospital discharge. For example, if a patient's last hospital discharge occurs on 2150-01-01, then the last possible date of death for the patient is 2151-01-01. If the individual died on or before 2151-01-01, and it was captured in either state or hospital death records, then the dod column will contain the deidentified date of death. If the individual survived for at least one year after their last hospital discharge, then the dod column will have a NULL value.

icu

The icu module contains data sourced from the clinical information system known as MetaVision (iMDSoft). MetaVision tables were denormalized to create a star schema where the icustays and d_items tables link to a set of data tables all suffixed with "events". Data documented in the icu module includes intravenous and fluid inputs (inputevents), ingredients for the aforementioned inputs (ingredientevents), patient outputs (outputevents), procedures (procedureevents), information documented as a date or time (datetimeevents), and other charted information (chartevents). All events tables contain a stay_id column allowing identification of the associated ICU patient in icustays, and an itemid column allowing identification of the concept documented in d_items. Additionally, the caregiver table contains caregiver_id, a deidentified integer representing the care provider who documented data into the system. All events tables (chartevents, datetimeevents, ingredientevents, inputevents, outputevents, procedureevents) have a caregiver_id column which links to the caregiver table.

The icu module contains a total of 94,458 ICU stays for 65,366 unique individuals as of MIMIC-IV v3.0. An ICU stay is defined as a contiguous sequence of transfers within a unit of the hospital classified as an ICU, and the icustays table is derived from the transfers table. During the creation of the icustays table, consecutive transfers within an ICU were merged into the same stay_id for analytical convenience, as these transfers are often bed number changes. Importantly, non-consecutive ICU stays remain as unique stay_id in the icustays table. In some cases, these could be considered the "same" ICU stay as the patient was transferred out for a planned procedure. In other cases, these are unanticipated readmissions to the ICU. As there was no systematically perfect method to differentiate these cases, we did not attempt to merge non-consecutive stay_id, and it is up to the investigator to appropriately handle these cases.

Usage Notes

The data described here are collected during routine clinical practice and reflect the idiosyncrasies of that practice. Implausible values may be present in the database as an artifact of the archival process.  Researchers should follow best practice guidelines when analyzing the data.

Documentation

Up to date documentation for MIMIC-IV is available on the MIMIC-IV website [4]. We have created an open source repository for the sharing of code and discussion of the database, referred to as the MIMIC Code Repository [5, 6]. The code repository provides a mechanism for shared discussion and analysis of all versions of MIMIC, including MIMIC-IV.

Linking MIMIC-IV to emergency department, note, and chest x-ray data

MIMIC-IV is linkable to other MIMIC projects published on PhysioNet. Where possible, we have prefixed the other projects with "MIMIC-IV" to make this clear such as MIMIC-IV-ED. Note that MIMIC-CXR is also linkable although it is not prefixed with MIMIC-IV. Free-text clinical notes are available in MIMIC-IV-Note [7], observations made in the emergency department are available in MIMIC-IV-ED [8], and chest x-rays in MIMIC-CXR [9].

Linking the other datasets to MIMIC-IV requires two steps. The first step is to match the data using subject_id, taking care to note that MIMIC-IV is a superset of other modules, and sampling biases may be introduced by the linking process. For example, MIMIC-CXR is only available between 2011 - 2016 for patients who were admitted to the emergency department, and this selection bias impacts the patient cohort. The second step involves aligning the dates. Since all modules are deidentified by the same shift, the time periods for measurements overlap. For example, if a patient is admitted to the hospital on 2105-01-01, discharged on 2105-01-03, and has an x-ray in MIMIC-CXR on 2105-01-02, then it is correct to assume the x-ray was taken while the patient was admitted to the hospital.

Patient composition

MIMIC-IV contains patients admitted to the emergency department and the intensive care unit. While patients admitted to the intensive care unit must have an associated hospitalization, patients may be admitted to the emergency department without being subsequently admitted to the hospital. As a result, the number of patients in MIMIC-IV is much higher than the number of unique patients with hospitalizations. As of MIMIC-IV v3.0 there are 364,627 unique patients, of whom 223,452 had at least one hospitalization (i.e. at least one record in the admissions table). The remaining 141,175 patients were only seen in the emergency department, which can be verified using the transfers table.

Release Notes

MIMIC-IV v3.1

MIMIC-IV v3.1 was released in October, 2024. This release fixed minor bugs raised by the community:

• The itemid values in the d_labitems and labevents tables had changed for a subset of laboratory measurements between v2.2 and v3.0. This change was not intentional. The tables have been updated, and the d_labitems and labevents itemid values have been verified to be consistent with v2.2.
• Two subject_id were present in various data tables but were not present in the patients table. These subjects have been removed from the data tables. Database constraints with a foreign key to the subject_id column in the patients table should now work correctly.

If upgrading from v3.0, note that only the following tables were modified (and thus require updating):

• d_labitems
• diagnoses_icd
• drgcodes
• labevents
• microbiologyevents
• omr
• transfers
• icustays

MIMIC-IV v3.0

MIMIC-IV v3.0 was released on July 23, 2024. Stays occurring between 2020 and 2022, inclusive, were added to the database. Out of hospital mortality is available for up to 1-year post hospital or ED discharge. The number of additional patients, admissions, and stays are highlighted by the increased row counts of their respective tables:

• patients: 364,627 (was 299,712 in v2.2)
• admissions: 546,028 (was 431,231 in v2.2)
• icustays: 94,458 (was 73,181 in v2.2)

Other changes include:

• Improved language data. The language column of admissions now provides a standardized primary language, if non-English, rather than "?" as before.
• Improved insurance data. The categories of the insurance column of admissions have been expanded to "Medicare", "Medicaid", "Private", "Self-pay", "No charge", and "Other". This change better aligns the field with other databases such as the National Inpatient Sample.

MIMIC-IV v2.2

MIMIC-IV v2.2 was released in January 2023. It added provider identifiers, imputed hadm_id for a number of rows in emar, and changed the subset of subject_id which are held out. Final row counts are available in the validation scripts published with the MIMIC Code Repository [6]. For clarity, after removal of the test set, the row counts are as follows:

• patients: 299,712 (was 315,460 in v2.0)
• admissions: 431,231 (was 454,324 in v2.0)
• icustays: 73,181 (was 76,943 in v2.0)

icu module

• caregiver
  • New table in v2.2. Contains one column: caregiver_id, a deidentified integer which uniquely represents a single caregiver or provider. These identifiers are sourced from the MetaVision ICU system. When present in a table, it indicates the user who documented the data into MetaVision. For example, the caregiver_id associated with a row indicating mechanical ventilation in the procedureevents table represents the user who documented the event, and not the provider who performed the procedure.
• chartevents, datetimeevents, ingredientevents, inputevents, outputevents, procedureevents
  • Added the caregiver_id column. This column is a deidentified integer representing the care provider who documented the data for the given row.

hosp module

• provider
  • New table in v2.2. Contains one column: provider_id, a deidentified string which uniquely represents a single caregiver or provider. These identifiers are sourced from the hospital wide EHR system, and used in a variety of contexts across tables in the module.
• admissions
  • New column: admit_provider_id, a deidentified string representing the provider who admitted the patient.
• emar
  • New column: enter_provider_id, a deidentified string representing the provider who entered the medication administration information into the database.
  • Fixed a bug where a subset of emar rows (713,117, ~2.5%) did not have an hadm_id even though they were associated with a given hospitalization. These rows occur outside of the administratively documented admission and discharge times for a hospitalization, but are still considered as administered during that hospitalization in the raw data.
• labevents, microbiologyevents, poe, prescriptions
  • New column: order_provider_id, a deidentified string representing the provider who ordered the corresponding event (e.g. the lab test in the case of labevents, or the medication in the case of prescriptions).

MIMIC-IV v2.1

MIMIC-IV v2.1 was released on November 16, 2022. It removed a subset of subject_id which will be retained internally as a test set. Future data releases will exclude these patients.

• patients - Removed 15,748 subject_id from the table
• admissions - Removed 23,093 hadm_id from the table.
• icustays - Removed 3,762 stay_id from the table.
• Other tables will have rows removed to reflect the removal of the aforementioned subject_id, hadm_id, and stay_id. Final row counts are available in the validation scripts published with the MIMIC Code Repository [6].

MIMIC-IV v2.0

MIMIC-IV v2.0 was released on June 12, 2022. It focused on expanding the data elements available for patients within MIMIC-IV v1.0. Additional data available includes out-of-hospital date of death, information from the online medical record system (which includes height and weight), and more detail for continuous infusions in the ICU.

Major changes

• The core module has been removed to simplify the schema. The admissions, patients, and transfers tables are now in the hosp module.
• Neonates have been removed from the dataset. Neonatal data will be released in a separate project with data from the neonatal intensive care unit.

icu module

• icustays
  • Around 700 stays (~1%) have changed due to the changes in the patients table.
• chartevents, d_items
  • The problem list from MetaVision has been added. All problems are documented with the same itemid now present in d_items: 220001. There are just over 1,000 unique problems. Most documented problems are related to the care plan for the patient and documented during nurse shift changes (either 7am or 7pm). Less frequently, the ongoing issues are documented here.
• ingredientevents
  • This is a new table associated with inputevents. Each intravenous administration tracked in inputevents is associated with a set of ingredients. These ingredients include water content, caloric information, and so on. The goal of the inputevents table is to support nutrition research and to provide a mechanism for estimating fluid input via summing all instances of the water ingredient. These ingredients have been separated from the inputevents table to simplify analysis and reduce the size of inputevents.
• inputevents
  • Removed a single column which contained only null values: cancelreason.
• procedureevents
  • Removed columns which contained only null values: totalamount, totalamountuom, cancelreason, comments_editedby, comments_canceledby, comments_date, secondaryordercategoryname.

hosp module

• admissions
  • Fixed an issue where hospitalizations were missing edregtime and edouttime when the patient was admitted via the ED (reported in #1247, thanks @MEladawi).
• patients
  • dod is now populated with out-of-hospital mortality from state death records. For patients admitted to the ICU, this change has increased capture of date of death from 8,223 records to 23,844 (i.e. we now have out-of-hospital mortality for an additional 15,621 ICU patients).
  • The mechanism for determining patients included in MIMIC was changed. For the most part this has resulted in an improvement, particularly regarding the logic for merging patients who had distinct medical record numbers. As a result of this change, most tables have had a change in the data content. Approximately 1% of stays were affected.
• transfers
  • Fixed a bug where the outtime for ED stays with no associated hadm_id (i.e. an ED stay where the individual was not admitted to the hospital) was incorrect. This resulted in all transfers rows with a NULL hadm_id having an apparent stay of minutes or less. The outtime column has now been corrected.
• labevents, d_labitems
  • The itemid for d_labitems has been changed for 43 items. These are extremely infrequently documented and each itemid has fewer than 100 observations in labevents. The exact itemid are provided in the changelog file CHANGELOG.txt.
  • Errors were found in the current values of loinc_code (reported in #938, thanks @Mauvila). In order to enable collaborative improvement, the loinc_code column has been removed, and will now be collaboratively developed in the MIMIC Code Repository. Initial values will be sourced from the hospital system.
  • A number of labs which previously had the value in the comments field now have the value in the value field (reported in #941, thanks @Mauvila). This change makes the labevents table more consistent with MIMIC-III, which had these values in the value field.
• microbiologyevents
  • New organisms, tests, specimens, and antibiotics have been added.
• omr
  • A new table has been added: omr. The source of this data is the Online Medical Record, and it contains miscellaneous information useful for understanding an individual's health. As of v2.0, the omr table has the following information: blood pressure, height, weight, body mass index, and Estimated Glomerular Filtration Rate (eGFR). These values are available from both inpatient and outpatient visits, and in many cases a "baseline" value from before a patient's hospitalization is available.
• prescriptions
  • The formulary_drug_cd table has been added back (was previously in MIMIC-III). This column has the same set of values as the product_code column of emar_detail.

MIMIC-IV v1.0

MIMIC-IV v1.0 was released March 16, 2021.

core

• admissions
  • A number (~1000, <1%) of erroneous hadm_id have been removed.
• patients
  • dod is now populated using the patient's deathtime from their latest hospitalization (reported in #71, thanks @jinjinzhou).
  • At the moment, out-of-hospital mortality is not captured by `dod`.
• transfers
  • Removed erroneous transfers included in the previous version.
  • transfer_id has been regenerated. transfer_id in MIMIC-IV v1.0 are not compatible with transfer_id from v0.4. We do not intend to change transfer_id when updating MIMIC-IV, but had to update it due to an error in its generation.
  • All hadm_id in transfers are also present in admissions and vice-versa (reported in #84, thanks @kokoko12305).

icu

• icustays
  • ICU stays were inappropriately assigned in the previous version due to an error in the preprocessing code. Previously, non-ICU ward transfers were included in the ICU stays, and certain ward stays were not treated as ICU stays (reported in #67, thanks @JHLiu7 and @stefanhgm). The assignment of stay_id has been regenerated.
  • The mapping between hospital transfers and ICU stays has been updated.
  • stay_id in MIMIC-IV v1.0 are not compatible with stay_id from v0.4. We do not intend to change stay_id when updating MIMIC-IV, but had to update it due to the error identified above.
• The change in icustays has re-assigned values to new stay_id, as a result all tables have had their content changed (due to a change in stay_id), but the structure is unchanged.

hosp

• hcpcsevents
  • Data has been added for a number of previously excluded hospitalizations.
  • The table now has a chartdate column, containing the date associated with the code. Every row is associated with a date.
• drgcodes
  • Data has been added for a number of previously excluded hospitalizations.
  • Duplicate DRG codes have been removed from the table.
  • Descriptions have been updated using the latest dictionaries made available from mass.gov and HCUP.
• diagnoses_icd, d_icd_diagnoses
  • Data has been added for a number of previously excluded hospitalizations (reported in #27, thanks @yugangjia).
  • The icd_code column is now trimmed and stored as a VARCHAR, i.e. codes no longer contain trailing whitespaces ('850 ' -> '850').
  • Missing ICD codes have been added to the dictionary. All ICD codes in the diagnoses_icd table have an associated reference in d_icd_diagnoses.
• labevents
  • The comments field has been updated, fixing a bug where comments longer than 4096 characters were truncated. Due to the deidentification, it's unlikely users will see much difference, as these comments will appear as ___.
• procedures_icd
  • Data has been added to procedures_icd for a number of previously excluded hospitalizations.
  • The table now has a chartdate column, containing the date associated with each billed procedure.
  • The icd_code column is now trimmed and stored as a VARCHAR, i.e. codes no longer contain trailing whitespaces ('850 ' -> '850').
  • Missing ICD codes have been added to the dictionary. All ICD codes in the procedures_icd table have an associated reference in d_icd_procedures.

v0.4

• d_micro
  • This table has been removed
• microbiologyevents
  • Added the column spec_type_desc, test_name, org_name, and ab_name columns
  • These columns contain the textual name of the organism/antibiotic/test/specimen
  • Added the comments column: this column contains information about the test, and in some cases (e.g. viral load tests), contains the result

v0.3

• Fixed a bug in the timing between hosp and icu

v0.2

• Updated demographics in the patient table
  • anchor_year -> anchor_year_group
  • anchor_year_shifted -> anchor_year
  • See the patients table in the MIMIC online documentation for detail on these columns
• transfers
  • Deleted the los column
• emar
  • mar_id -> emar_id
  • emar_id is now a composite of subject_id and emar_seq, and has form “subject_id-emar_seq”
  • emar_seq column - a monotonically increasing integer starting with the first eMAR administration
  • Added poe_id and pharmacy_id columns for linking to those tables
• emar_detail
  • mar_id -> emar_id (changed as above)
  • Deleted the mar_detail_id column
• hcpcsevents
  • ticket_id_seq -> seq_num
• labevents
  • Many previously NULL values are now populated - these were removed originally due to deidentification
  • Added the comments column. This contains deidentified free-text comments with labs. PHI is replaced with three underscores (___). If an entire comment is ___, then the entire comment was scrubbed.
• microbiologyevents
  • stay_id column removed
  • spec_id -> micro_specimen_id
• Added the poe and poe_detail tables
  • Documentation of provider orders for various treatments and other aspects of patient management
• Added the prescriptions table
  • Documentation of medicine prescriptions via the provider order interface
• Added the pharmacy table
  • Detailed information regarding prescriptions provided by the pharmacy including formulary dose, route, frequency, dose, and so on.
• inputevents
  • Fixed an error in the calculation of the amount column
• icustays
  • Re-derived stay_id - the new stay_id are distinct from the previous version.

Ethics

The collection of patient information and creation of the research resource was reviewed by the Institutional Review Board at the Beth Israel Deaconess Medical Center, who granted a waiver of informed consent and approved the data sharing initiative.

Acknowledgements

We would like to thank the Beth Israel Deaconess Medical Center for their continued support of the MIMIC project. In particular we would like to thank Carolyn Conti, Alvin Gayles, Larry Markson, Ayad Shammout, Lu Shen, and Manu Tandon for their assistance. This work was supported by the National Institute of Biomedical Imaging and Bioengineering (NIBIB) under NIH grant number R01EB030362.

Conflicts of Interest

None to declare.

References

1. Henry, J., Pylypchuk, Y., Searcy T. & Patel V. (May 2016). Adoption of Electronic Health Record Systems among U.S. Non-Federal Acute Care Hospitals: 2008-2015. ONC Data Brief, no.35. Office of the National Coordinator for Health Information Technology: Washington DC.+
2. Halevy, A., Norvig, P., & Pereira, F. (2009). The unreasonable effectiveness of data. IEEE Intelligent Systems, 24(2), 8-12.
3. Johnson, A. E., Pollard, T. J., Shen, L., Lehman, L.H., Feng, M., Ghassemi, M., ... & Mark, R. G. (2016). MIMIC-III, a freely accessible critical care database. Scientific data, 3(1), 1-9.
4. MIMIC Online Documentation. https://mimic.mit.edu
5. Johnson AE, Stone DJ, Celi LA, Pollard TJ. The MIMIC Code Repository: enabling reproducibility in critical care research. Journal of the American Medical Informatics Association. 2018 Jan;25(1):32-9.
6. Alistair Johnson, Tom Pollard, Jim Blundell, Brian Gow, erinhong, Nicolas Paris, et al. MIT-LCP/mimic-code: MIMIC Code v2.1.1. Zenodo; 2021. https://doi.org/10.5281/zenodo.821871
7. Johnson, A., Pollard, T., Horng, S., Celi, L. A., & Mark, R. (2023). MIMIC-IV-Note: Deidentified free-text clinical notes (version 2.2). PhysioNet. https://doi.org/10.13026/1n74-ne17.
8. Johnson, A., Bulgarelli, L., Pollard, T., Celi, L. A., Mark, R., & Horng, S. (2023). MIMIC-IV-ED (version 2.2). PhysioNet. https://doi.org/10.13026/5ntk-km72.
9. Johnson, A., Pollard, T., Mark, R., Berkowitz, S., & Horng, S. (2019). MIMIC-CXR Database (version 2.0.0). PhysioNet. https://doi.org/10.13026/C2JT1Q.