MIMIC-IV - Medical Information Mart for Intensive Care

One academic medical center's two-decade intensive care record, opened up as structured data instead of filing cabinets. MIMIC-IV - Medical Information Mart for Intensive Care assembles the de-identified electronic health record of patients treated at Beth Israel Deaconess Medical Center in Boston between 2008 and 2022, curated by the MIT Laboratory for Computational Physiology and distributed under the PhysioNet banner. Version 3.1 (published October 2024) covers 364,627 unique individuals. The hosp module spans hospital-wide ground for 546,028 hospitalizations serving 223,452 of those people: admissions and transfers, laboratory results, microbiology cultures, prescriptions, electronic medication administrations, pharmacy orders, ICD-coded diagnoses and procedures, billing and service assignments. The icu module follows 65,366 individuals through 94,458 ICU stays with hour-level charted vitals, fluid inputs and outputs drawn from the bedside clinical system. An ed module adds emergency department visits with triage scores, and companion note, cxr and ecg modules carry discharge and radiology free text, 377,110 chest X-ray images across 227,835 studies, and electrocardiogram machine measurements. It scores 10/10 in our catalog - a mark only 145 of the 1,744 datasets we researched reach.

One patient-to-stay chain, exactly as it lands in your warehouse. Values shown are the documented example values carried on each field definition:

subject_id                        : 10000032
hadm_id                           : 22595853       stay_id : 31933343
admission_type                    : EW EMER.
admittime / dischtime             : shifted timestamps inside a 2100-2200 window
anchor_year / anchor_age          : populated      hospital_expire_flag : 0 or 1
itemid                            : populated      valuenum / valueuom : result plus unit
charttime / storetime             : event time versus documentation time
starttime / endtime               : drug administration window
icd_code / icd_version            : populated (9 or 10)
insurance / language /
marital_status / race             : populated per hospitalization

Read it as a join kit. subject_id keys the person, hadm_id keys the hospitalization, stay_id keys the ICU or ED stay, and itemid joins every event row back to the lab and charted-variable dictionaries. The anchor-year pair restores calendar arithmetic on top of the shifted timestamps, so a timeline computed in 2153 still means exactly what it says. Sepsis windows, dosing intervals and readmission gaps all hang off that identifier spine.

Thirteen documented field groups anchor the dictionary, verified against the source's published schema documentation during research. Module-specific detail tables - microbiology antibiograms, pharmacy fill lines, ED triage and med-reconciliation rows, radiology report text and imaging metadata - ride on the same identifiers and ship under additional fields scoped at sample request rather than promised in every cut.

Geography - deliberately deep rather than broad: a single US academic medical center, Beth Israel Deaconess Medical Center in Boston, Massachusetts. Every lab draw, charted vital, medication dose, note and image comes from one institution's real operations, which is precisely why the record supports internal-validity claims most multi-site aggregates cannot.

Temporal - 2008 through 2022 for the core record, with the chest X-ray subset spanning 2011 to 2016. Raw calendar dates are replaced during de-identification: patient events land inside a 2100-2200 window, and each patient carries an anchor year and anchor age so year-level aging, seasonality and multi-admission spacing stay computable without exposing anyone's true dates.

Granularity - four nested levels: patient, then hospitalization, then ICU or ED stay, then the individual event row underneath - one row per lab result, per charted observation, per medication administration. Imaging sits at DICOM study level, paired with its radiology report.

API, files, or your warehouse. Daily, weekly, or hourly.

Hourly suits live evaluation rigs - a model under test reading the same stay-level panels your researchers do, refreshed fast enough that a demo dashboard never shows stale vitals. Daily suits production pipelines: cohort features recomputed each morning while the previous day's runs reconcile cleanly against the last. Weekly suits research rhythms - benchmark reruns, paper revisions, quarterly cohort refreshes where stability matters more than latency.

Every cut reuses the identical field names and the same identifier spine, so snapshots concatenate into longitudinal panels without remapping a single column. Request the modules you need - hospital-wide, ICU, emergency, notes, imaging - and the sample arrives cut to them, with the field dictionary and coverage profile attached.

• Sepsis, deterioration and outcome modeling - hour-level charted vitals joined to lab results and medication administrations give onset-prediction models a defensible ground truth; mapped to the ml-model-training use case.
• Clinical language models - discharge summaries and radiology reports aligned to structured outcomes turn free-text benchmarks into measurable ones, from summarization fidelity to diagnosis-code prediction.
• Imaging-and-report pairs - 377,110 chest X-ray studies keyed to their radiology text support multimodal retrieval, report-generation and label-noise studies at a scale few institutional partners will share.
• Operations and health-services research - admission types, transfer chains, service assignments and payer mix support length-of-stay, readmission and capacity studies grounded in actual hospital operations.
• Pharmacovigilance and dosing studies - start-and-stop timestamps on drug administrations plus fluid inputs and outputs expose treatment timing at minute resolution.
• Benchmark lineage - the record sits behind a large share of published critical-care ML work, so a new model's numbers land on a comparable footing; see the citation-grade research use case.

Data scientists and ML engineers get the deepest single-site clinical panel in the catalog: event-level grain, stable identifiers and a dictionary that documents itself, the difference between a publishable cohort and a weekend of schema archaeology. Journalists and academics get citable, versioned numbers on what intensive care actually looks like - stays, mortality flags, payer mix - without waiting on an institutional partnership. Investors and quants read care patterns, utilization and documentation intensity as ground truth beneath health-technology theses pitched on operational efficiency. Product teams building healthcare software get a realistic test corpus whose quirks - missingness, shift charts, unit mismatches - match production reality far better than synthetic fixtures ever will.

Three things worth knowing upfront. First, scope is depth over breadth: one Boston academic medical center, so treat external validity as a hypothesis your own multi-site data tests rather than something this record promises. Second, time is de-identified - absolute calendar dates are shifted into a 2100-2200 window, so analyses needing true wall-clock dates map through the anchor-year pair instead; plan for that join before you build. Third, volume needs scoping: the imaging and waveform companions are measured in hundreds of thousands of studies, so name the modalities and year windows you want when requesting a sample and the cut arrives sized to the question rather than to the archive.