The MIMIC database

What is MIMIC?

The MIMIC database is a North American database containing data from over 50,000 patients admitted to intensive care at the Beth Israel Deaconess Medical Center (Boston, USA). It is developed and maintained by the MIT Lab for Computational Physiology.

It is one of the most widely used intensive care databases in the scientific literature, thanks to its public access and the richness of its data: demographics, diagnoses, lab results, medication prescriptions, clinical notes, physiological signals, and more.

Available versions

MIMIC comes in several versions:

• MIMIC-III (2016): data from 2001 to 2012, approximately 46,000 patients, MIMIC-specific data schema.
• MIMIC-IV (2023): most recent version. Data from 2008 to 2019, approximately 65,000 patients, modernized schema.

Both versions can be converted to the OMOP CDM format using open source ETLs: MIT-LCP/mimic-omop for MIMIC-III, OHDSI/MIMIC for MIMIC-IV. A demo database of MIMIC-IV OMOP with 100 patients is freely available (see below).

Demo databases (open access)

Demo databases are publicly available. They contain anonymized data from 100 patients and require no registration.

You can download them directly:

• MIMIC-III Demo — MIMIC data schema
• MIMIC-IV OMOP Demo — OMOP CDM v5.4 data schema

These demo databases are ideal for discovering the data structure and practicing SQL queries before accessing the full data.

Accessing the full data

To access the full databases, you need to complete a few steps.

Useful links:

• PhysioNet registration
• Credentialing request
• CITI Course instructions
• Certificate submission
• MIMIC-IV page

Downloading the files

Once access is granted, the Files section at the bottom of the MIMIC-IV page lets you download the files from your browser. You can also fetch everything at once with the wget command shown on that same page: open a terminal (the Terminal app on Mac), replace YOUR_USERNAME with your PhysioNet username, then type your password when prompted.

cd ~/Downloads
wget -r -N -c -np --user YOUR_USERNAME --ask-password https://physionet.org/files/mimiciv/3.1/

The downloaded files (about 8 GB) are compressed CSVs, stored in the hosp/ and icu/ folders. With wget, they end up in physionet.org/files/mimiciv/3.1/.

Database structure

Native MIMIC schema

The native MIMIC-IV schema is organized around several modules:

• hosp: hospital data (admissions, diagnoses, prescriptions, lab results)
• icu: intensive care-specific data (physiological measurements, severity scores, inputs/outputs)
• ed: emergency department data
• note: free-text clinical notes

Key tables include patients, admissions, diagnoses_icd, labevents, prescriptions, chartevents, and more.

OMOP schema

The OMOP version of MIMIC-IV follows the OMOP CDM v5.4 standard. The data is reorganized into standardized tables: person, visit_occurrence, condition_occurrence, measurement, drug_exposure, etc. You can explore the full schema on our interactive explorer.

The advantage of the OMOP format is that you can use the same SQL queries on any database converted to OMOP, whether it’s MIMIC, a French hospital warehouse, or any other compliant database.

Loading MIMIC-IV into Linkr

Once you have downloaded MIMIC-IV, you can explore it in Linkr within minutes, without installing a database server.

Converting the files to Parquet

PhysioNet provides MIMIC-IV as compressed CSV files (.csv.gz): one file per table, stored in the hosp/ and icu/ folders. Linkr reads Parquet files, a column-based storage format that is much more compact and faster to query. The files therefore need to be converted once.

Open a terminal, then copy and paste the block below. If you did not use wget, adapt the first line: the path to the folder that contains hosp/ and icu/.

# Go to the downloaded folder (the one that contains hosp/ and icu/)
cd ~/Downloads/physionet.org/files/mimiciv/3.1

# Install DuckDB, the tool that does the conversion (only once)
curl https://install.duckdb.org | sh

# Convert each table to Parquet, into a new parquet/ folder
for f in hosp/*.csv.gz icu/*.csv.gz; do
  out="parquet/${f%.csv.gz}.parquet"
  mkdir -p "$(dirname "$out")"
  echo "Converting $f"
  ~/.duckdb/cli/latest/duckdb -c "COPY (FROM read_csv('$f', delim = ',', sample_size = -1)) TO '$out'"
done

# Go to the downloaded folder (the one that contains hosp\ and icu\)
cd "$HOME\Downloads\physionet.org\files\mimiciv\3.1"

# Install DuckDB, the tool that does the conversion (only once)
winget install DuckDB.cli --accept-source-agreements --accept-package-agreements
$env:Path = [Environment]::GetEnvironmentVariable("Path", "User") + ";" + [Environment]::GetEnvironmentVariable("Path", "Machine")

# Convert each table to Parquet, into a new parquet\ folder
foreach ($f in Get-ChildItem hosp, icu -Filter *.csv.gz) {
  $src = "$($f.Directory.Name)/$($f.Name)"
  $out = "parquet/" + ($src -replace '\.csv\.gz$', '.parquet')
  New-Item -ItemType Directory -Force -Path (Split-Path $out) | Out-Null
  Write-Host "Converting $src"
  duckdb -c "COPY (FROM read_csv('$src', delim = ',', sample_size = -1)) TO '$out'"
}

You get a parquet/ folder with the same layout: parquet/hosp/admissions.parquet, parquet/icu/icustays.parquet, and so on.

Creating the database in Linkr

Linkr detects the tables automatically: the hosp/ and icu/ subfolders become the hosp and icu SQL schemas, so you can write queries such as SELECT * FROM hosp.admissions. The data stays on your machine or your server: Linkr does not send it anywhere.

The database’s Statistics tab then gives a first overview: number of patients, hospitalizations and unit stays, sex distribution, lengths of stay and admission timeline.

Once the database is loaded, you can browse the concept dictionary, select cohorts by criteria, and display each patient’s individual record on a timeline. The page Explore a health data warehouse walks through each of these screens, using the MIMIC-IV Demo.