Health Care Technology Data Providers · Head-to-head
MIMIC-IV - Medical Information Mart for Intensive Care vs PhysioNet - Research Resource for Complex Physiologic Signals
Which health care technology data providers data fits your job: MIMIC-IV - Medical Information Mart for Intensive Care, or PhysioNet - Research Resource for Complex Physiologic Signals. API, files, or your warehouse. Daily, weekly, or hourly.
MIMIC-IV - Medical Information Mart for Intensive Care
PhysioNet - Research Resource for Complex Physiologic Signals
What each contains
Pick by fit, not by loyalty.
| MIMIC-IV - Medical Information Mart for Intensive Care | PhysioNet - Research Resource for Complex Physiologic Signals | |
|---|---|---|
| Patient identity | `subject_id` - integer, one row per patient, repeating across admissions and event tables | `subject_id` - string linking records within a hosted project |
| Episode or recording key | `hadm_id` (hospitalization, range 2000000-2999999) and `stay_id` (ICU or ED stay) | `record_name` indexing WFDB `.hea`/`.dat` files; `study_id` mapping DICOM studies to subjects |
| Observed measurement | `valuenum` / `valueuom` - numeric result plus unit of measure for labs and charted variables | `signal` / `channel` - named waveform channels such as ECG leads, arterial pressure, airway flow |
| Time axis | `charttime` / `storetime` event and documentation times; `starttime` / `endtime` for administrations | `sample_rate` - sampling frequency in Hz declared per channel (VitalDB runs at hundreds of Hz) |
| Vocabulary lookup | `itemid` joining `d_labitems` and `d_items`; `icd_code` / `icd_version` (9 or 10) for diagnoses and procedures | None centralized - each project declares its own channels in header files |
| Demographics and outcomes | `insurance` / `language` / `marital_status` / `race`; `hospital_expire_flag`; `admission_type` across nine classes; `anchor_year` / `anchor_age` | Not standardized across hosted projects |
| Documentation status | 13 fields, definitions verified against official schema documentation | 6 fields, definitions inferred from project conventions |
What each does better
MIMIC-IV - Medical Information Mart for Intensive Care
Hour-level clinical resolution nothing else in the pool reaches. The icu module holds 94,458 ICU stays for 65,366 individuals with charted vitals, laboratory events, medication administrations and fluid outputs stamped to the hour; the hosp module widens the frame to 546,028 hospitalizations for 223,452 individuals. Reading a sepsis timeline, a titration curve or a 72-hour fluid balance starts here.
Outcome structure ready for modeling. Every hospitalization carries hospital_expire_flag, ICD-coded diagnoses and procedures with the version stated (icd_code/icd_version at 9 or 10), admission type across nine urgency classes, and payer, language, marital status and race demographics - labels most signal collections leave to the researcher to construct.
Privacy engineering already applied. De-identification follows HIPAA Safe Harbor: true dates are shifted into 2100-2200 and replaced by anchor_year/anchor_age, so cohort math works year by year without exposing a birth date - done once at the corpus level instead of per project.
One spine across six modalities. Structured labs, discharge notes, 377,110 chest X-rays with their radiology reports and an ECG module all resolve back to the same subject_id, so a multimodal model trains on one join surface.
PhysioNet - Research Resource for Complex Physiologic Signals
Breadth of physiology one hospital cannot supply. Several hundred curated resources span device-generated waveforms, imaging, clinical text, annotations, software and ML models. VitalDB alone contributes high-fidelity multi-parameter vital signs from surgical patients in Seoul, with channels sampled at hundreds of Hz - signal density the EHR tables' hourly grain deliberately trades away.
Geographic and institutional spread. Contributors and collections circle the globe: VitalDB from South Korea, MIMIC-BR bringing 31,789 anonymized Brazilian hospital and ICU admissions, SCRIPT CarpeDiem and 1970s-era ECG archives sharing one catalog, with the 190,000-plus registered users spanning more than 180 countries.
Versioning, benchmarks and tooling. Every resource publishes as a versioned component with discovery DOIs and per-version DOIs; annual PhysioNet Challenges turn portions of the shelf into public benchmarks, cited across more than 49,000 articles since 2001. Open-source waveform software and tutorials ride alongside, so signal processing starts from working code rather than a blank notebook.
The verdict
Verdict: sample both, pick by fit - they tie at 10 out of 10, so the question decides, not the leaderboard.
Pick MIMIC-IV - Medical Information Mart for Intensive Care when the unit of analysis is a patient-stay: deterioration prediction on hour-level vitals, medication response from administration timestamps, discharge-note NLP, ICD-keyed outcome cohorts, or multimodal work joining labs, notes and chest X-rays on one patient spine.
Pick PhysioNet - Research Resource for Complex Physiologic Signals when the unit of analysis is a recording: ECG denoising or beat classification, multi-parameter vital-sign fusion at hundreds of Hz, challenge-baseline reproduction, or teaching signal processing against several hundred varied collections.
Three quick tests settle most cases. Diagnosis codes or mortality flags? Only the corpus carries them. Waveforms from outside one Boston hospital? Only the shelf reaches them. One study moving from device signal to clinical outcome is the both-of-them case - start from the data scientists use cases page.
Sample both, pick by fit. See MIMIC-IV - Medical Information Mart for Intensive Care · See PhysioNet - Research Resource for Complex Physiologic Signals
Fair questions
Is MIMIC-IV - Medical Information Mart for Intensive Care better than PhysioNet - Research Resource for Complex Physiologic Signals?
Different instruments, tied on craft - both score 10/10, a mark only 145 of 1,744 cataloged datasets reach. The corpus wins when the question follows a patient through a stay: 94,458 ICU stays with hour-level vitals, labs, medications, notes and 377,110 chest X-rays from one hospital between 2008 and 2022. PhysioNet wins when the question follows a signal: several hundred collections including VitalDB and MIMIC-BR, versioned with DOIs and cited in more than 49,000 articles. Many teams keep both.
Which dataset covers more patients?
It depends which direction you count. Within one record, the corpus is unmatched in this pairing: 364,627 unique individuals, 546,028 hospitalizations and 94,458 ICU stays from a single medical center. Across records, PhysioNet's shelf aggregates far more - several hundred resources whose flagships run from thousands of VitalDB recordings to MIMIC-BR's 31,789 Brazilian admissions, drawn on by more than 190,000 registered users.
Do the two datasets overlap anywhere?
Literally, yes: MIMIC-IV is one of the collections hosted inside the PhysioNet repository, alongside MIMIC-CXR and MIMIC-BR. The field dictionaries echo that nesting - `subject_id` and `study_id` appear on both sides, and both records descend from the same MIT Laboratory for Computational Physiology stewardship. Beyond the MIMIC family, the shelf's VitalDB holdings and ECG archives cover ground the corpus never touches.
Which one should a data scientist sample first?
Sample both, pick by fit - the unit of analysis decides. Modeling patient outcomes, sepsis timelines or discharge-note language starts with [MIMIC-IV](/datasets/health-care-technology/mimic-iv-medical-information-mart-for-intensive-care). Modeling waveforms - ECG beats, vital-sign fusion at hundreds of Hz, challenge baselines - starts with [PhysioNet](/datasets/health-care-technology/physionet-research-resource-for-complex-physiologic-signals). Samples arrive cut to your modules, signal types and year windows either way.
Can Datadory deliver both datasets together?
Yes. Either record arrives alone or both arrive merged onto one calendar, aligned on subject identity and time grain so the join is done before it reaches you - delivered daily, weekly, or hourly, your call. Name the modules, signal types and windows when you request the sample and it lands pre-cut, with field definitions and coverage profiles attached. Or take both in one feed.