MIMIC-IV Dataset: Modular EHR Schema, MIMIC-III Differences, and Access

MIMIC-IV (Medical Information Mart for Intensive Care IV) is a large-scale, retrospective, de-identified clinical research database developed by the MIT Laboratory for Computational Physiology (LCP) in partnership with Beth Israel Deaconess Medical Center (BIDMC) in Boston, Massachusetts (Johnson et al., Scientific Data, 2023). Hosted on PhysioNet, MIMIC-IV serves as a primary benchmark dataset for clinical machine learning, critical care epidemiology, pharmacovigilance, and algorithmic fairness evaluations across electronic health record (EHR) workflows.

What Does MIMIC-IV Stand For and What Data Does It Contain?

MIMIC-IV stands for Medical Information Mart for Intensive Care IV. Unlike earlier critical care registries restricted strictly to patients inside an intensive care unit, MIMIC-IV captures comprehensive longitudinal hospital-wide EHR trajectories alongside high-resolution bedside ICU telemetry for patients admitted to the BIDMC emergency department or intensive care units.

In release v2.2 (covering admissions from 2008 to 2019), the core database encompasses 299,712 unique patients (`subject_id`), 431,231 hospital admissions (`hadm_id`), and 73,181 ICU stays (`stay_id`); subsequent releases (v3.0 and v3.1) extend data collection through 2022 to encompass more than 364,000 unique patients, 546,000 hospitalizations, and 94,000 ICU stays.

How Does the MIMIC-IV Modular Schema Work?

To preserve data provenance across distinct hospital software systems, MIMIC-IV organizes relational tables into two primary core modules (`hosp` and `icu`) linked by three hierarchical keys—subject_id (patient), hadm_id (hospital admission), and stay_id (ICU or ED stay)—and connects seamlessly to four external modality modules on PhysioNet:

Module	Source Clinical System	Core Relational Tables & Data Elements
hosp (Hospital-Wide EHR)	Hospital billing, laboratory, microbiology, and pharmacy EHR systems	patients (anchor age, sex, date of death), admissions, transfers, labevents & d_labitems (LOINC-mapped blood/fluid assays), microbiologyevents (cultures and antibiotic susceptibilities), prescriptions, emar & emar_detail (barcode-scanned electronic Medication Administration Records), diagnoses_icd, and procedures_icd (ICD-9-CM and ICD-10-CM codes).
icu (Intensive Care Unit)	iMDsoft MetaVision bedside clinical information system	icustays, chartevents (vital signs, ventilator settings, Glasgow Coma Scale), inputevents (continuous IV infusions and boluses), outputevents (urine and drain outputs), procedureevents, datetimeevents, and d_items.
MIMIC-IV-ED	Emergency Department information system	Emergency department stays (`edstays`), triage acuity scores (`triage`), periodic ED vital signs (`vitalsign`), medication reconciliation (`medrecon`), and pyxis dispensing (`pyxis`).
MIMIC-IV-Note & MIMIC-CXR	Clinical documentation & PACS imaging archives	De-identified free-text discharge summaries and radiology reports (`MIMIC-IV-Note`) plus over 377,000 chest radiographs in DICOM and JPEG formats (`MIMIC-CXR` / `MIMIC-CXR-JPG`) linkable via subject_id and timestamps.

What Is the Difference Between MIMIC-III and MIMIC-IV?

Investigators migrating analytical pipelines or SQL cohort definitions from MIMIC-III (v1.4) to MIMIC-IV must account for five structural changes:

1. Separation of hosp and icu Schemas: MIMIC-III stored all tables in a single monolithic schema restricted to ICU patients. MIMIC-IV separates hospital-wide tables (`hosp`) from bedside ICU tables (`icu`) and includes hospital admissions for patients who visited the ED or inpatient wards without necessarily entering the ICU during every hospitalization.
2. Single ICU Bedside System (MetaVision Only): MIMIC-III merged two incompatible ICU clinical information systems—Philips CareVue (2001–2008) and iMDsoft MetaVision (2008–2012)—requiring researchers to reconcile duplicate item IDs across `inputevents_cv` and `inputevents_mv`. MIMIC-IV begins in 2008 and sources `icu` data exclusively from MetaVision, providing a single unified `inputevents` table.
3. Barcode eMAR Medication Tracking: Whereas MIMIC-III captured inpatient non-ICU medications primarily as physician order ranges (`prescriptions`), MIMIC-IV adds `emar` and `emar_detail`, recording exact timestamped barcode administrations at the patient bedside.
4. Dual ICD-9 and ICD-10 Coding: Because MIMIC-IV spans the October 2015 U.S. clinical coding transition, `diagnoses_icd` and `procedures_icd` include an `icd_version` column (`9` or `10`), unlike MIMIC-III which contained exclusively ICD-9-CM codes.
5. Expanded Mortality Tracking: MIMIC-IV links in-hospital mortality (`admissions.deathtime`) with out-of-hospital mortality from state vital records (`patients.dod`) up to one year post-discharge.

What Standards Govern MIMIC-IV De-Identification and Access?

MIMIC-IV is shared under Institutional Review Board (IRB) approvals from BIDMC and MIT with a waiver of informed consent because the dataset is thoroughly de-identified under the HIPAA Privacy Rule Safe Harbor standard (45 CFR § 164.514(b)(2)). To remove all 18 HIPAA identifiers while preserving longitudinal clinical utility, the de-identification pipeline applies three transformations:

• Randomized Patient-Level Date Shifting: All timestamps for a given `subject_id` are shifted by a uniform random offset (measured in days) into a future century window (between the years 2100 and 2200). The offset preserves exact within-patient chronological intervals (down to the minute), time of day, day of the week, and approximate seasonality. To allow researchers to identify the true era of care without exposing calendar dates, the `patients` table provides `anchor_year` (the shifted year) and `anchor_year_group` (a 3-year real calendar interval such as 2014 - 2016).
• Age Top-Coding Above 89: In compliance with HIPAA Safe Harbor rules, any patient whose age at `anchor_year` exceeds 89 years is assigned an `anchor_age` of 91 regardless of their true age above 89.
• Free-Text PHI Scrubbing: Clinical notes in `MIMIC-IV-Note` undergo automated rule-based and neural-network data masking that replaces patient names, clinician names, hospital unit names, telephone numbers, and explicit dates with triple underscores (___).

To obtain the dataset, researchers follow the PhysioNet credentialed access workflow: create a PhysioNet profile, submit an institutional reference, upload a completion report for the CITI Program Data or Specimens Only Research training module, and execute the PhysioNet Credentialed Health Data Use Agreement (DUA). Once credentialed, investigators can download the `.csv.gz` tables directly or query MIMIC-IV in the cloud via Google Cloud BigQuery (`physionet-data.mimiciv_hosp` and `physionet-data.mimiciv_icu`) and AWS Open Data.

Frequently Asked Questions

What does MIMIC-IV stand for and what is it used for in clinical research?

MIMIC-IV stands for Medical Information Mart for Intensive Care IV. Developed by the MIT Laboratory for Computational Physiology in collaboration with Beth Israel Deaconess Medical Center (BIDMC), it is a large, de-identified relational electronic health record dataset covering emergency department, hospital-wide, and intensive care unit admissions used globally for clinical machine learning, observational epidemiology, and reproducible health informatics research.

What is the difference between MIMIC-III and MIMIC-IV?

Whereas MIMIC-III contained ICU admissions from 2001 to 2012 in a single monolithic schema merging legacy CareVue and MetaVision bedside systems, MIMIC-IV covers contemporary admissions (2008–2019 in v2.2; extended through 2022 in v3.x) and splits data into modular provenance schemas: a hospital-wide EHR module (hosp) including non-ICU inpatient admissions and electronic Medication Administration Records (eMAR), and an ICU module (icu) sourced exclusively from the MetaVision clinical information system, plus linked ED, clinical note, and chest X-ray modules.

Is MIMIC-IV publicly available and how do researchers access it?

MIMIC-IV is freely available to researchers worldwide through PhysioNet under a credentialed-access governance model rather than unrestricted public download. To gain access, an investigator must register a PhysioNet account, complete the CITI Program ‘Data or Specimens Only Research’ human subjects training course, verify institutional affiliation, and sign the PhysioNet Credentialed Health Data Use Agreement (DUA).