MIMIC-III Critical Care Database: 26 Relational Tables, CareVue vs. MetaVision, NOTEEVENTS, and HIPAA Date-Shifting Explained

MIMIC-III (Medical Information Mart for Intensive Care III) is a freely accessible, de-identified relational critical care database developed by the MIT Laboratory for Computational Physiology and published by Alistair E. W. Johnson, Tom J. Pollard, Lu Shen, et al. in 2016 (Scientific Data, 3, 160035). Encompassing 53,423 distinct hospital admissions for 46,520 unique patients (38,597 adults aged 16 years or older and 7,875 neonates) admitted to critical care units at Beth Israel Deaconess Medical Center (BIDMC) in Boston, Massachusetts between 2001 and 2012, MIMIC-III v1.4 remains the most widely cited benchmark dataset in clinical time-series modeling, SOFA score sepsis phenotyping, and biomedical natural language processing alongside its modular successor MIMIC-IV and imaging benchmarks such as the CheXpert dataset.

What Is MIMIC-III? Cohort Scope and the 26 Relational Tables

Unlike disease-specific clinical registries that capture only low-frequency summary variables at admission and discharge, MIMIC-III integrates high-resolution bedside physiological waveforms/vital signs, laboratory results, medication administration records, fluid balance logs, ICD-9-CM billing codes, free-text clinical notes, and out-of-hospital mortality records from the Social Security Administration Death Master File.

In MIMIC-III v1.4, all clinical data are distributed as 26 comma-separated value (CSV) relational tables linked by three hierarchical integer primary/foreign keys:

• SUBJECT_ID: Identifies a unique individual patient across the entire database (46,520 distinct patients).
• HADM_ID: Identifies a single hospital admission (58,976 total hospital admissions across all patients).
• ICUSTAY_ID: Identifies a contiguous intensive care unit stay derived by grouping ICU transfers within 24 hours (61,532 total ICU stays across MICU, SICU, CSRU, CCU, TSICU, and NICU).

Table Category	MIMIC-III Table Names	Primary Identifiers & Lookup Dictionary	Core Clinical Data Captured
Patient & Stay Tracking (5 tables)	PATIENTS, ADMISSIONS, ICUSTAYS, CALLOUT, TRANSFERS	SUBJECT_ID, HADM_ID, ICUSTAY_ID	Demographics, DOB/DOD, admission/discharge timestamps, physical ward/bed transitions, ICU discharge readiness
ICU Bedside Charted Events (5 tables)	CHARTEVENTS, DATETIMEEVENTS, INPUTEVENTS_CV, INPUTEVENTS_MV, OUTPUTEVENTS	ITEMID → joined to D_ITEMS; CGID → joined to CAREGIVERS	330+ million vital signs, ventilator settings, GCS scores, IV fluid/vasopressor infusions, and urine/drain outputs
Hospital & Lab Data (4 tables)	LABEVENTS, MICROBIOLOGYEVENTS, PRESCRIPTIONS, PROCEDUREEVENTS_MV	ITEMID → joined to D_LABITEMS (for labs) or D_ITEMS	Hospital-wide chemistry/hematology labs, blood/urine culture sensitivities, medication orders, MetaVision bedside procedures
Unstructured Clinical Text (1 table)	NOTEEVENTS	ROW_ID, SUBJECT_ID, HADM_ID, CATEGORY, TEXT	2,083,180 de-identified free-text discharge summaries, nursing notes, physician progress notes, ECG, and radiology reports
Administrative & Billing (5 tables)	CPTEVENTS, DIAGNOSES_ICD, DRGCODES, PROCEDURES_ICD, SERVICES	ICD9_CODE → joined to D_ICD_DIAGNOSES and D_ICD_PROCEDURES	Hospital discharge ICD-9-CM diagnoses and procedures, Diagnosis-Related Groups (DRGs), CPT procedures, clinical service assignments
Dictionary & Metadata (6 tables)	CAREGIVERS, D_CPT, D_ICD_DIAGNOSES, D_ICD_PROCEDURES, D_ITEMS, D_LABITEMS	ITEMID, CGID, ICD9_CODE	Human-readable labels, LOINC mappings (in D_LABITEMS), source database (`dbsource`), and caregiver roles (MD, RN, RT)

The CareVue (2001–2008) vs. MetaVision (2008–2012) EHR Transition

The single most common methodological pitfall when querying MIMIC-III is failing to account for the ICU clinical information system migration that occurred at Beth Israel Deaconess Medical Center in 2008:

• Philips CareVue Era (2001–2008): Recorded bedside nursing observations, vital signs, and infusions using ITEMID values below 200000 (typically 1 to 49999). Intravenous fluid and continuous medication infusions from this era are stored in INPUTEVENTS_CV, which records point-in-time volume boluses/infusions at a single CHARTTIME without explicit start and stop timestamps.
• iMDsoft MetaVision Era (2008–2012): Assigned completely new ITEMID codes starting at 220000 and above. Intravenous infusions are stored in a separate table, INPUTEVENTS_MV, which captures rate-based intervals with explicit STARTTIME and ENDTIME columns, patient weights, and order cancellation flags (STATUSDESCRIPTION).

Because CHARTEVENTS and D_ITEMS retain both code sets side by side (distinguished by the DBSOURCE column as carevue or metavision), querying a single ITEMID for heart rate (for example, only CareVue 211 while omitting MetaVision 220045) silently drops half the cohort. Reproducible MIMIC-III pipelines always rely on the community-validated SQL concept definitions in the MIT-LCP mimic-code repository to harmonize CareVue and MetaVision ITEMID concepts and unit conversions.

HIPAA Safe Harbor De-Identification and Date-Shifting Architecture

To comply with the Health Insurance Portability and Accountability Act (HIPAA) Safe Harbor standard (45 CFR § 164.514(b)) while preserving longitudinal clinical utility, MIMIC-III applies three structural de-identification transformations:

1. Consistent Per-Patient Future Date Shifting: All calendar dates (admission, discharge, birth, death, lab draw, and chart timestamps) for a given SUBJECT_ID are shifted into the future—falling between the years 2100 and 2200—by a random patient-specific offset. Critically, the offset is constant across all hospital admissions for the same patient, and it preserves the exact time of day, day of the week, and approximate season. Consequently, researchers can accurately compute ICU length of stay, hourly vasopressor titration intervals, and 30-day readmission rates, but cannot correlate a patient’s shifted year (for example, 2142) with real-world calendar years unless comparing broad CareVue vs. MetaVision eras.
2. Age Masking for Patients Over 89 Years Old: Because HIPAA Safe Harbor classifies exact ages above 89 as protected health information, any patient aged 90 or older at their first admission had their date of birth shifted so that their calculated age at first admission appears as ≥ 300 years (with a median shifted age of 300). Analysis scripts must recode ages > 89 (or >= 300) to a capped value (such as 90 or 91.4) before training regression or neural models to prevent extreme numeric outliers.
3. Free-Text PHI Surrogates in NOTEEVENTS: Automated rule-based and dictionary lookup de-identification algorithms (derived from Neamatullah et al., 2008) scrubbed patient names, attending physician names, hospital locations, telephone numbers, and exact dates from the 2.08 million notes in NOTEEVENTS, replacing each span with bracketed PHI tags such as [**Known lastname 1234**], [**Hospital1 18**], or [**2145-03-14**].

How Is MIMIC-III Used in Clinical Research and Machine Learning?

Since 2016, MIMIC-III has underpinned thousands of peer-reviewed studies across three main research tracks:

• Standardized Clinical Time-Series Benchmarks (Harutyunyan et al., 2019): Established canonical train/validation/test splits over 17 physiological variables for four multi-task benchmark problems: (1) in-hospital mortality prediction from the first 48 hours of an ICU stay, (2) physiologic decompensation prediction within the next 24 hours, (3) length-of-stay forecasting, and (4) 25-class acute care phenotype classification.
• Clinical Natural Language Processing (`NOTEEVENTS`): Used to pretrain domain-specific clinical language models—most notably ClinicalBERT (Alsentzer et al., 2019) and BlueBERT—and to benchmark multi-label automated ICD-9 medical coding (such as the CAML 50-code and full-code benchmarks using discharge summaries). Note that when predicting in-hospital mortality or readmission from NOTEEVENTS, researchers must filter out rows where CATEGORY = 'Discharge summary' (which are written at the end of the stay and contain explicit mortality/disposition outcomes) and remove notes marked with ISERROR = 1.
• Target Trial Emulation and Sepsis-3 Reinforcement Learning: Used to derive the Third International Consensus Definitions for Sepsis and Septic Shock (SOFA score) and to evaluate offline reinforcement learning policies for dynamic intravenous fluid and vasopressor dosing.

Frequently Asked Questions

How do researchers gain access to MIMIC-III?

Although MIMIC-III is de-identified and exempt from individual patient consent, it is a restricted-access dataset hosted on PhysioNet. Investigators must create a PhysioNet account, complete human subjects research training (such as CITI “Data or Specimens Only Research”), verify their institutional identity, and sign the PhysioNet Credentialed Health Data Use Agreement prohibiting re-identification. For step-by-step instructions, see PhysioNet Credentialed Access: How to Get MIMIC and Other Restricted Data.

Should new clinical AI studies use MIMIC-III or MIMIC-IV?

New clinical epidemiology and EHR modeling studies should generally use MIMIC-IV, which covers contemporary admissions (2008–2019+), uses native MetaVision/hospital schemas (`hosp` and `icu` modules), includes ICD-10-CM codes alongside ICD-9-CM, and links to MIMIC-CXR chest radiographs and MIMIC-IV-ED emergency department records. MIMIC-III remains essential when reproducing legacy baselines (such as Harutyunyan et al., 2019) or evaluating models trained on its specific NOTEEVENTS corpus.

Why are some rows in MIMIC-III CHARTEVENTS and NOTEEVENTS flagged with errors?

Clinicians at the bedside occasionally enter a measurement or note under the wrong patient or correct a typo after saving. In CHARTEVENTS, erroneous rows are flagged with ERROR = 1 (in MetaVision) or STOPPED = 'D/C', and in NOTEEVENTS erroneous notes have ISERROR = '1'. Standard preprocessing pipelines always exclude these flagged rows prior to feature extraction.