What's new in MIMIC-IV?

Many users will be familiar with MIMIC-III, the predecessor of MIMIC-IV. A number of improvements have been made, including simplifying the structure, adding new data elements, and improving the usability of previous data elements.

Table-wise changes are summarized below.

MIMIC-III table	MIMIC-IV module	Removed columns	Additional columns	Notes
patients	hosp	row_id, dob, dod_ssn, dod_hosp, expire_flag	anchor_age, anchor_year, anchor_year_group	DOD no longer available outside of the hospital.
admissions	hosp	row_id, religion, has_chartevents_data		Values changed for many columns (e.g. language, ethnicity).
transfers	hosp	row_id, icustay_id, dbsource, prev_careunit, curr_careunit, prev_wardid, curr_wardid, los	transfer_id, careunit	Schema is simplified in MIMIC-IV.
d_hcpcs	hosp	N/A, was d_cpt		New table.
d_icd_diagnoses	hosp	row_id, icd9_code, short_title	icd_code, icd_version	Now contains ICD-9 and ICD-10 codes.
d_icd_procedures	hosp	row_id, icd9_code, short_title	icd_code, icd_version	Now contains ICD-9 and ICD-10 codes.
d_labitems	hosp	row_id		Many additional itemid added.
diagnoses_icd	hosp	row_id, icd9_code	icd_code, icd_version	Now contains ICD-9 and ICD-10 codes.
drgcodes	hosp	row_id
emar	hosp	N/A, new table.
emar_detail	hosp	N/A, new table.
hcpcsevents	hosp	N/A, was cptevents.
labevents	hosp	row_id	labevent_id, specimen_id, storetime, ref_range_lower, ref_range_upper, priority, comments
microbiologyevents	hosp	row_id	microevent_id, micro_specimen_id, test_seq, storedate, storetime, test_itemid, test_name, quantity, comments
pharmacy	hosp	N/A, new table.
poe	hosp	N/A, new table.
poe_detail	hosp	N/A, new table.
prescriptions	hosp	row_id, startdate, enddate, drug_name_poe, drug_name_generic, formulary_drug_cd	starttime, stoptime, form_rx, doses_per_24_hrs
procedures_icd	hosp	row_id, icd9_code	icd_code, icd_version	Now contains ICD-9 and ICD-10 codes.
services	hosp	row_id
d_items	icu	row_id, dbsource, conceptid	lownormalvalue, highnormalvalue
chartevents	icu	row_id, icustay_id, cgid, error, resultstatus, stopped	stay_id
datetimeevents	icu	row_id, icustay_id, cgid, error, resultstatus, stopped	stay_id
icustays	icu	row_id, icustay_id	stay_id
inputevents	icu	row_id, icustay_id, cgid, error	stay_id	Was inputevents_mv in MIMIC-III.
outputevents	icu	row_id, icustay_id, cgid, error	stay_id
procedureevents	icu	row_id, icustay_id, cgid, comments_editedby, comments_canceledby	stay_id, patientweight, totalamount, totalamountuom, originalamount, originalrate	Removed columns had no data in MIMIC-III

Structure

The structure of MIMIC-IV is necessarily different than MIMIC-III. In MIMIC-III, the set of tables were given as one large set, with no obvious differentiation between them. In MIMIC-IV, we explicitly state the source database of each table. Not only does this clarify the data provenance, but it answers many questions regarding data coverage. For example, as the CHARTEVENTS table is sourced from the ICU clinical information system, it will only provide data for patients while they are in an ICU. Conversely, the LABEVENTS table is sourced from the hospital database, and consequently contains information for a patient’s entire hospital stay.

Contemporary

MIMIC-IV contains data from 2008 - 2019 (inclusive). Biomarkers which have been more recently introduced will be available.

CareVue is no more

As the update covers the years 2008 - 2019, the CareVue clinical information system is no longer relevant, as it was not used during that time period. The implications are:

• All itemid in d_items with a value less than 220000 are no longer relevant.
• The inputevents_cv table has been removed. The inputevents_mv table was renamed inputevents. The structure is otherwise unchanged.
• The procedureevents_mv table has been renamed procedureevents.

icustay_id is now stay_id

Eventually, stays across different areas of the hospital will be indexed by a unique stay_id, such that a stay in the emergency department, ICU, and operating room will all be distinct and referred to by the same identifier. In preparation for this change, icustay_id has been renamed stay_id.

Years are included

The date-shift strategy in MIMIC has changed. Instead of releasing the day of the week and the season, we have released the approximate year of patient admission. This allows studying patients over time as care practices change.

Audit trails are removed

MIMIC-III contained a number of rows associated with auditing clinical documentation. These rows were marked as erroneous in various ways (error = 1, statusdescription = ‘Rewritten’). These rows have been removed in MIMIC-IV.

Chest x-ray data

Imaging data is also an entirely new addition to MIMIC. The MIMIC-CXR database is publicly available. Notably, the subject_id identifier used in the MIMIC-CXR database is consistent with the subject_id used in MIMIC-IV. Therefore, all chest x-rays in MIMIC-CXR are linkable to patient stays in MIMIC-IV.

Table-wise improvements over MIMIC-III

A number of enhancements have been made to tables which may be familiar to you from MIMIC-III. Entirely new tables have also been added.

Hospital data

emar and emar_detail

Two entirely new tables are made available, sourced from the relatively newly installed electronic Medicine Administration Record (eMAR) system. Bedside staff will scan barcodes for each individual formulary unit of a medication when administering it. This allows for a granular, high resolution record of when a medication was given.

labevents

• Reference ranges are now available.
• A specimen identifier (specimen_id) allows users to group all measurements made for a single specimen (e.g. all blood gas measurements from the same sample of blood).
• A priority column indicates the priority level of the laboratory measure.

microbiologyevents

• Now contains the name of the test performed.

prescriptions

• Instead of startdate and enddate, prescriptions now has starttime and stoptime.
  • This means all prescriptions now have the date and time of start/stop
  • In an internal assessment, only 10 prescriptions were missing the start hour, and 1650 prescriptions were missing the stop hour (there are over 17 million rows in this table).
  • We cannot guarantee the start time is the first instance of patient administration (as these are prescriptions), but the added resolution should help in research studies.
• drug_name_generic, drug_name_poe, and formulary_drug_cd have been removed.
  • drug_name_poe, when not null, was always equal to drug.
  • drug is the displayed drug name in the EHR, and is more reliable than drug_name_generic.
  • formulary_drug_cd was an internal ontology that did not provide additional information over drug.
• New columns!
  • pharmacy_id - to link to the pharmacy table which has additional information about the prescription
  • form_rx.
  • doses_per_24_hrs provides the number of doses per 24 hours prescribed by this row.