MIMIC-IV-Ext MIMIC Note Extract: A Large-Scale Structured Dataset Extracted from MIMIC-IV-Note Discharge Summaries

@article{PhysioNet-mimic-iv-ext-note-extract-1.0,
  author = {Kuntzer, Jonas and Schwethelm, Kristian and Ketzer, Conrad and Wende, Eva and Schlager, Moritz and Rueckert, Daniel and Kaissis, Georgios and Jungmann, Friederike and Hager, Paul},
  title = {{MIMIC-IV-Ext MIMIC Note Extract: A Large-Scale Structured Dataset Extracted from MIMIC-IV-Note Discharge Summaries}},
  journal = {{PhysioNet}},
  year = {2026},
  month = oct,
  note = {Version 1.0},
  doi = {10.13026/cq50-rd11},
  url = {https://doi.org/10.13026/cq50-rd11}
}

@article{pollard_physionet_2026, title = {{PhysioNet} as a global platform for biomedical research}, volume = {1}, issn = {3005-0693}, url = {https://doi.org/10.1038/s44360-026-00096-z}, doi = {10.1038/s44360-026-00096-z}, abstract = {Twenty-five years after its first public repository was established, the PhysioNet resource for biomedical and clinical data has enabled tens of thousands of research papers. Further efforts are ongoing to advance secure data sharing, community engagement and scalable research infrastructure.}, number = {8}, journal = {Nature Health}, author = {Pollard, Tom and Moody, Benjamin E. and Lehman, Li-wei H. and Gow, Brian J. and Fernandes, Chrystinne and Xie, Chen and Johnson, Alistair and Mark, Roger G. and Heldt, Thomas}, month = aug, year = {2026}, pages = {792--795},}

Abstract

MIMIC is one of the most widely used public hospital datasets for machine learning on electronic health records [3], but its structured tables answer many clinical questions only through coded surrogates, or not at all. For example, a patient's past medical history, documented allergies, and pertinent diagnostic findings are only available in MIMIC's discharge summaries. Studies that need this information must therefore extract it from unstructured text, a labor-intensive process that is repeated independently across projects.

We release MIMIC-Note-Extract (MNE), a large-scale dataset containing data from all 16 sections of 300,720 MIMIC-IV-Note v2.2 discharge summaries. We use regex rules and large language models (LLMs) to transform these sections into structured tables while preserving their clinical content. Validation by three medical doctors confirms extraction accuracy above 92% across sections.

Background

The structured tables of MIMIC-IV [1] are missing much of the narrative context of patient care, which appear only in unstructured MIMIC-IV-Note [2] discharge summaries. This leads to a fragmented ecosystem of individual preprocessing where, in the absence of a standardized pipeline, each project must build its own scripts to extract the specific variables needed for its use case. These isolated preprocessing pipelines are rarely shared or unified, resulting in duplicated engineering effort and inconsistent extraction quality across the community. But even if such an effort were made, the resulting datasets are typically tailored so closely to individual projects that they lack the generality needed to support broader research questions. At the core of this lies a methodological limitation: most existing projects focus on information extraction. Instead, we posit that the community requires information conversion; preserving the underlying information of the raw data while transforming it into an easily queryable, structured format.

We introduce MIMIC-Note-Extract (MNE), a structured dataset derived from 300,720 MIMIC-IV-Note discharge summaries. We employ a hybrid processing framework that pairs deterministic parsing with schema-guided LLM generation to translate 16 major clinical sections of the discharge summary into structured tables. By providing structured representations for 13 of these sections, spanning pre-admission history to detailed physical examinations, MNE makes the information directly accessible for downstream use.

Methods

Cohort construction

We derive the cohort from the 331,793 raw discharge summaries in MIMIC-IV-Note v2.2 [2]. Quality filtering removes non-discharge note types, formatting outliers such as specific external dictation system templates, and reports with disordered or severely incomplete clinical sections. The final dataset comprises 300,720 hospital admissions from 137,402 patients.

Splitting into Sections

We identified 24 distinct clinical sections that appear in consistent relative order across MIMIC-IV-Note discharge summaries. Eight of these are entirely removed by upstream de-identification, for example, Name, Admission Date, and Date of Birth are dropped, leaving 16 sections.

Splitting is complicated by partially redacted headers. For example, the header "Discharge ___" may indicate either "Discharge Condition" or "Discharge Disposition". Our deterministic splitting algorithm maps common header variants to canonical section names and exploits the consistent relative ordering of sections to infer headers that remain ambiguous after redaction [8]. A residual 5,461 ambiguous cases could not be resolved automatically and were handled manually.

Information modeling

Tabular target schemas were developed iteratively in close consultation with clinicians. The difficulty is the structural and content diversity within individual sections: a Pertinent Results section ranges from a single-line negative culture report to multiple pages of microbiology panels and longitudinal imaging findings. Schemas therefore had to capture essential clinical depth without introducing hyper-specific fields that remain empty across most admissions, and had to stay concise enough for human verification. We analyzed source texts for recurring patterns to refine field definitions.

Three sections (history of present illness, brief hospital course, discharge instructions) vary too much between cases to admit a reliable field decomposition and are preserved verbatim.

Adaptive extraction

We use deterministic rule-based extraction (regular expressions and string matching) for simple cases and LLMs for complex narratives that require contextual understanding:

• Rule-based: gender, services, discharge disposition
• Hybrid: chief complaint, allergies, major procedures, discharge diagnosis, and discharge condition
• LLM-only: past medical history, medications on admission, physical examination, pertinent results, discharge medications

LLM extraction used Gemini 2.5 Flash [5] and Gemini 3.1 Flash-Lite [6]. Complete prompts and generation settings are available in the code repository [8]. We provide the few-shot example files used for LLM extraction and the curated allergy standardization map in the extraction_resources/ directory. These resources contain sensitive clinical text and are therefore excluded from the code repository.

Privacy and de-identification

We used Gemini through the Google Cloud Vertex AI API with zero data retention. This follows PhysioNet’s requirements for using MIMIC data with LLMs and online services [7].

The discharge summaries were broadly de-identified by the original developers [2], with redactions applied conservatively to avoid retaining personal identifying information (PII), which also removed some clinically relevant information. We deliberately did not attempt to recover PII and only normalized MIMIC's redaction artifacts to [REDACTED].

Quality validation

Three medical doctors reviewed extractions side by side against the source text. Rule-based extractions were excluded from manual review because they are deterministic and auditable from the published code. Annotation was performed at the item level, which distinguishes a single incorrect field from an entirely incorrect sample.

For each section we sampled at random 171 admissions, split into 121 single-annotated and 50 dual-annotated cases. Overlap cases were independently assigned to two of the three medical doctors, which allows inter-annotator agreement to be measured while maximizing coverage. We report Gwet's AC1 [4] as a chance-corrected agreement metric. The annotators completed 2,210 case-level tasks comprising 13,744 item-level decisions.

Item-level accuracy ranges from 92.2% to 99.8% across sections, as shown in Table 1.

Data Description

The dataset covers 300,720 hospital admissions from 137,402 distinct patients, one discharge summary per admission. It comprises 16 CSV files holding 19.3 million rows in total.

Sixteen discharge summary sections are provided as verbatim free text. 13 of those are additionally extracted into 14 structured tables, where the physical examination section splits into a vitals table and a body systems table. The three text-only sections are history of present illness, brief hospital course, and discharge instructions.

Structure and Keys

Every table links to the hub table admissions through hadm_id. The admissions table contains subject_id, hadm_id, and note_id from the discharge summaries in MIMIC-IV-Note.

Six tables hold at most one row per admission and are keyed on hadm_id alone. These are admissions, discharge_sections, gender, services, discharge_disposition, and discharge_condition. Eight tables hold multiple rows per admission and are keyed on (hadm_id, seq_num), where seq_num preserves the order in which entries appear in the source section. The two physical examination tables are keyed on (hadm_id, timepoint) and (hadm_id, timepoint, system).

Files

Hubs

• admissions.csv: the link table containing subject_id, hadm_id, and note_id
• discharge_sections.csv: the verbatim text of all 16 sections after structural parsing and before any structured extraction

Section Tables

• gender.csv: recorded sex (M or F)
• services.csv: the service at discharge (enum over 20 services)
• discharge_disposition.csv: the discharge destination (enum over 4 categories).
• discharge_condition.csv: discharge condition, covering mental status, functional activity, and further clinical status fields
• allergies.csv: lists allergies
• chief_complaint.csv: lists chief complaints and context
• discharge_diagnosis.csv: lists clinician-written primary and secondary discharge diagnoses
• major_procedures.csv: lists major procedures
• past_medical_history.csv: structured past medical history (prior conditions, family, substance, and functional history)
• medications_on_admission.csv: admission medications including name, presentation form, dose, route, frequency, strength, duration, and PRN and hold flags
• discharge_medications.csv: discharge medications structured like admission medications
• physical_exam_vitals.csv: lists vital signs from physical examinations including timepoint (admission, discharge, interim, or unknown)
• physical_exam_systems.csv: lists examined systems as an enum over 11 body systems, where status records normal, abnormal, or not assessed and abnormal_findings lists the documented findings. Includes the timepoint as well.
• pertinent_results.csv: identified result blocks, each carrying a content type enum, the line range locating the block in the source section, per-type enum arrays, and a topic label for clinical notes

Data Dictionary

This dictionary describes all fields in the 16 CSV files of MIMIC-Note-Extract. Types correspond to the PostgreSQL loading schema: integer, real, text, varchar, boolean, enum, arrays, and jsonb.

admissions.csv

discharge_sections.csv

gender.csv

Recorded sex (M or F).

services.csv

The service at discharge (enum over 20 services).

discharge_disposition.csv

The patient's discharge destination.

discharge_condition.csv

The patient's discharge condition, covering mental status, functional activity, and further clinical status fields.

allergies.csv

List of a patient's allergies.

chief_complaint.csv

List of a patient's chief complaints and the admission context.

discharge_diagnosis.csv

List of clinician-written primary and secondary discharge diagnoses.

major_procedures.csv

List of a patient's major procedures.

past_medical_history.csv

List of a patient's past medical history (prior conditions, family, substance, and functional history).

medications_on_admission.csv

List of a patient's admission medications including name, presentation form, dose, route, frequency, strength, duration, and PRN and hold flags.

discharge_medications.csv

List of a patient's discharge medications (structured like admission medications).

Medication controlled vocabularies

Medication strength JSON

Illustrative single-strength object:

{"type":"SINGLE","numerator_values":[10.0],"numerator_units":["MG"],"denominator_value":1.0,"denominator_unit":"TABLET"}

Medication frequency JSON

A frequency with type equal to [OTHER] contains no additional fields.

Illustrative frequency objects:

{"type":"times_per_period","times":2,"period_unit":"DAY"}
{"type":"every_period","value":8,"unit":"HOUR"}
{"type":"specific_time","time":"BEDTIME"}
{"type":"specific_day","days":["MONDAY","WEDNESDAY","FRIDAY"]}
{"type":"[OTHER]"}

physical_exam_vitals.csv

Lists a patient's vital signs from physical examinations.

physical_exam_systems.csv

List of a patient's examined systems as an enum over 11 body systems, where status records normal, abnormal, or not assessed and abnormal_findings lists the documented findings. Includes the timepoint of the examination as well.

pertinent_results.csv

List of a patient's pertinent results (lab tests, imaging studies, ECGs, etc.).

Usage Notes

Raw CSVs

The extracted discharge summary sections are provided as gzipped CSV files.

Note: discharge_sections.csv contains verbatim clinical text with embedded newlines within quoted fields. CSV readers must therefore support quoted multiline fields.

PostgreSQL

We recommend loading the files into PostgreSQL using the scripts in our code repository [8], which supply typed columns, enums, primary keys, and foreign keys to admissions. Extract the CSV files, then run create.sql to define the schema and load.sql to load the data.

Croissant

We also provide Croissant 1.1 metadata [9], which describes the files, column types, keys, foreign keys to admissions, and controlled vocabularies. The metadata can be used directly with the published gzip files.

Intended use cases

The database supports a broad range of applications by making clinical information from discharge summaries available in structured tables that can be linked to the other MIMIC-IV tables. Use case examples include:

• Cohort identification: Selecting admissions based on documented clinical histories, findings, or treatments.
• Medication reconciliation: Comparing pre-admission and discharge medications with inpatient prescription records.
• Allergy-related prescribing research: Studying associations between reported allergy labels, antibiotic prescribing, and clinical outcomes.
• Risk adjustment: Incorporating documented medical history and functional status into outcome analyses.
• Clinical decision-making benchmarks: Constructing patient cases for developing and evaluating diagnostic models and agents.

Potential limitations

The database has limitations arising from its extraction methods and source data:

• Extraction errors: Section splitting and rule-based or LLM-based extraction may have introduced errors. Clinician review of sampled extractions supports the extractions' quality but cannot establish correctness for every record.
• Incomplete representation: Predefined schemas do not capture all information in the original summaries, and some information was intentionally omitted during schema design. The schemas were iteratively refined with clinicians to retain clinically relevant content. Although unlikely, important details may still have been overlooked. Missing entries should not be interpreted as evidence that a condition or finding was absent.
• Cohort selection: Excluding incomplete or unusually formatted summaries may have introduced selection bias.
• Source-data limitations: The database inherits MIMIC-IV-Note's single-center coverage, incomplete or inaccurate clinical documentation, and information loss from de-identification.

Release Notes

Version 1.0.0

Initial release

Ethics

The dataset is a derivative dataset of MIMIC-IV and thus no new patient data was collected. The ethics approval of the dataset follows from that of the parent MIMIC-IV-Note dataset.

Conflicts of Interest

No conflicts of interest.

References

1. Johnson, A., Bulgarelli, L., Pollard, T., Horng, S., Celi, L. A., & Mark, R. (2023). MIMIC-IV (version 2.2). PhysioNet. RRID:SCR_007345. https://doi.org/10.13026/6mm1-ek67
2. Johnson, A., Pollard, T., Horng, S., Celi, L. A., & Mark, R. (2023). MIMIC-IV-Note: Deidentified free-text clinical notes (version 2.2). PhysioNet. RRID:SCR_007345. https://doi.org/10.13026/1n74-ne17
3. Zhang, X., Kong, W., Shi, R., Sun, L., Xu, M., & Gong, L. (2025). Data-driven trends in critical care informatics: a bibliometric analysis of global collaborations using the MIMIC database (2004-2024). Computers in Biology and Medicine, 195, 110670. https://doi.org/10.1016/j.compbiomed.2025.110670
4. Gwet, K. L. (2008). Computing inter-rater reliability and its variance in the presence of high agreement. British Journal of Mathematical and Statistical Psychology, 61(1), 29-48. https://doi.org/10.1348/000711006X126600
5. Google DeepMind. Gemini 2.5 Flash, 2025. URL https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-2-5-Flash-Model-Card.pdf Accessed: 16.06.26
6. Google DeepMind. Gemini 3.1 Flash Lite, 2026. URL https://deepmind.google/models/model-cards/gemini-3-1-flash-lite/ Accessed: 16.06.26
7. Use of MIMIC data with large language models and online services. https://physionet.org/news/post/llm-responsible-use/ Accessed: 27.07.26
8. mimic-note-extract. https://github.com/kschwethelm/mimic-note-extract/ Accessed: 20.10.26
9. MLCommons. Croissant format specification. https://docs.mlcommons.org/croissant/docs/croissant-spec.html Accessed: 28.07.26