AIR·MS Data Modalities

Mount Sinai Data Warehouse (MSDW) OMOP De-Identified (De-ID) and Protected Health Information (PHI)

The MSDW dataset leverages the OMOP Common Data Model. The data is comprised of clinical data extracted from Mount Sinai’s Epic Caboodle database and other ancillary systems. ​​​​​​​We offer both the identified (PHI) and de-identified (De-ID) versions of this data.

We offer both an identifiable and a de-identified version of the MSDW dataset in AIR·MS

To access the data, follow our “Getting Started” guide here: https://labs.icahn.mssm.edu/minervalab/air%E2%80%A7ms-getting-started/ 

MSDW OMOP identifiable (PHI) with extended attributes ​​​​​​​Current Status

Schema: CDMPHI
Data snapshot from 09/10/2025
Unique patients: 12,521,239

Note: Some of the standard OMOP tables contain extension fields (starting with the prefix ‘XTN’) which contain data outside of the OMOP standard data model. Many of these XTN attributes are based on data derived directly from EPIC (i.e. codes used in EPIC rather than the standardized OMOP codes), or attributes not currently contained in the OMOP standard. 

MSDW OMOP de-identified (de-id) ​​​​​​​Current Status

Schema: CDMDEID
Data snapshot from 10/20/2025
Unique patients: 12,008,581

Pathology Metadata (PHI)

The Pathology metadata aids researchers in the field of Computational Pathology. Researchers are able to query the metadata in combination with other linked data modalities to build a patient cohort, subsequently apply quantitative methods for the analysis of digital microscopy slides and relating the resulting statistical descriptors to patient outcomes. We are also working on making the digital slides available to researchers on Minerva HPC.

To access the data, follow our “Getting Started” guide here: https://labs.icahn.mssm.edu/minervalab/air%E2%80%A7ms-getting-started/ 

Current Status

Data Source: Powerpath
Schema: CDMPATHOLOGY
Data snapshot from 01/08/2025
Unique patients: 3,528,719

Note: Pathology reports are now available in ​​​​​​​​​​​​​​​​​​​​​AIR·MS and can be found in table ACC_RESULTS. The reports are broken up into sections (clinical history, final diagnosis, SNOMED coding, etc.) that can be identified by column PATH_RPT_HEADING_NAME. If you are looking for the full report, you can combine all records for with the same ACCESSION_2_ID into a single output. The column containing the free-text (ACC_RESULTS_FINDING) has been enabled with SAP HANA full-text search capabilities. Examples of how to use HANA full-text search are available in this tutorial notebook.

The following tables and attributes are available in AIR·MS:

ACCESSION

Mount Sinai Million Health Discoveries Program

The current lack of diversity in genomic research data is hindering what we can learn about health and potential treatments in our global population. By enhancing the diversity of people participating in genomic research, we can advance our knowledge and discovery of human genetics for all populations. To that end, The Charles Bronfman Institute for Personalized Medicine is spearheading the effort to carry out the genetic sequencing of one million Mount Sinai patients within the next five years. This initiative, one of the largest such sequencing projects of its kind, will integrate health and research data at Mount Sinai to promote discoveries that will directly benefit our patient population. Access to the BioMe Biobank and Mount Sinai Million Biobank on HPC can be requested via the CBIPM Data and Specimen Inquiry Form. AIR·MS now features radiology metadata extracted from the Mount Sinai IRW 2.0 XNAT system (via an MSDW data pipeline). This data set is comprised of detailed DICOM (Digital Imaging and Communications in Medicine) tags associated with the medical images. These tags provide essential metadata, including patient information, imaging parameters, equipment details, and procedural context, ensuring a comprehensive understanding of each radiological study. By integrating this metadata, we enable researchers to gain deeper insights into the imaging data, facilitating advanced analyses and fostering innovations in medical imaging research.

To access the data, follow our “Getting Started” guide here: https://labs.icahn.mssm.edu/minervalab/air%E2%80%A7ms-getting-started/ 

The following tables and attributes are available in AIR·MS:

Schema: CDMMSM
Data snapshot from 06/09/2025
Unique patients: 279,885

Schema: CDMMSMDEID
Data snapshot from 06/09/2025
Unique patients: 279,885

Electrocardiogram (ECG) Data (PHI)​

Electrocardiogram data, derived from Mount Sinai’s Cardiology Information System, is now available in AIR·MS.

To access the data, follow our “Getting Started” guide here: https://labs.icahn.mssm.edu/minervalab/air%E2%80%A7ms-getting-started/ 

Data Source: GE HealthCare MUSE Cardiology Information System

Schema: CDMECG
Data snapshot from: 04/10/2021
Unique patients: 1,961,254

The following tables and attributes are available in AIR·MS:

Number of records: 9,275,130

Number of records: 9,168,266

Number of records: 73,631,055

Number of records: 9,654,325

Intensive Care Unit (ICU) Data (PHI)

The Mount Sinai ICU Datamart is the world’s first ICU data platform designed to simultaneously support research, quality improvement, and operational initiatives. It is built on a common data model that standardizes critical care medical concepts, enabling consistent interpretation and integration of data across diverse ICU settings.The Mount Sinai ICU Datamart harmonizes and delivers high-fidelity information from all Mount Sinai adult ICUs with highly granular data available from 2011 onward and refreshed weekly. Beyond serving as a comprehensive data resource, it also tracks the evolution of the health system’s critical care landscape, capturing changes in unit specialties, geographic distribution, and the addition of new ICUs. By transforming the ICU’s inherently rich data environment into a standardized, dynamic, and accessible platform, the Mount Sinai ICU Datamart empowers clinicians, researchers, and administrators to advance data-driven care, operational excellence, and clinical discovery.

To access the data, follow our “Getting Started” guide here: https://labs.icahn.mssm.edu/minervalab/air%E2%80%A7ms-getting-started/ 

Schema: CDMICU
Number of patients: 103,974
Number of hospital admissions: 128,599
Number of ICU stays: 152,154
Number of clinical events and observations: 553,095,763

Number of records: 128,600

Number of records: 289,789,014

Number of records: 17

Number of records: 18,030

Number of records: 263,306,749

Number of records: 1,858,853

Number of records: 63

Number of records: 104,046

Number of records: 525,267

Number of records: 839,542

Number of records: 152,154

Number of records: 289,789,014

Number of records: 30

Number of records: 29

Number of records: 5

Number of records: 47

Number of records: 336

Number of records: 189

Number of records: 38

Number of records: 8

Number of records: 85

Number of records: 298

Number of records: 176

Number of records: 11

Radiology Image Data (PHI and De-ID)

AIR·MS provides a new radiology dataset organized in a ground-up, OMOP-aligned medical imaging common data model. The currently implemented model represents DICOM studies and series and links those records to OMOP patients and concepts. The identifiable schema also stores study- and series-level DICOM attributes that are not promoted to dedicated columns while preserving nested sequence structure, repeated tags, arrays, coded values, units, and typed values.

The identifiable model contains source identifiers and other protected health information (PHI). The de-identified model removes direct identifiers, converts linked identifiers to de-identified keys, shifts patient dates and datetimes, and transforms free-text values to remove PHI.

To access the data, follow the AIR·MS “Getting Started” guide.

Current Status

Identifiable schema: CDMRADIOLOGY
De-identified schema: CDMRADIOLOGYDEID
Distinct persons with a populated study access URI: 546,136 in CDMRADIOLOGY; 541,243 in CDMRADIOLOGYDEID
Model definition: September 4, 2026

Record Counts

Model Structure

The model follows the native DICOM hierarchy:

IMAGING_STUDY
├── IMAGING_STUDY_ATTRIBUTE (PHI only)
└── IMAGING_SERIES
    └── IMAGING_SERIES_ATTRIBUTE (PHI only)

Tables

Key Relationships

• IMAGING_STUDY.PERSON_ID links an imaging study to the OMOP person record.
• IMAGING_SERIES.IMAGING_STUDY_ID links each series to its parent study.
• In CDMRADIOLOGY, each attribute table links to its owning study or series and uses a self-referencing parent attribute ID to represent nested DICOM sequences.
• OMOP concept fields provide standardized mappings where available; a value of 0 indicates that no mapping is available.

De-Identification Behavior

The implemented CDMRADIOLOGYDEID schema contains IMAGING_STUDY and IMAGING_SERIES. Attribute tables are available only in the identifiable CDMRADIOLOGY schema. The De-ID model applies the following rules to its implemented tables:

• Patient, study, and series identifiers are de-identified and stored as char(64) values.
• Study, series, acquisition, content, and instance-creation date/datetime values are shifted on a patient-specific basis.
• Free-text study descriptions, series descriptions, institution names, and department names are algorithmically transformed rather than retaining verbatim PHI-bearing text.
• Direct patient identifiers and source DICOM identifiers are excluded, including accession number, Study Instance UID, Series Instance UID, DICOM patient ID, patient name components, birth date, sex, and address.
• Source-system ETL keys and ETL audit fields are excluded from the De-ID model.
• Study- and series-level attribute tables, including their nested DICOM values, are not currently implemented in the De-ID schema.

Nested DICOM Tags, Sequences, and Arrays

The two PHI attribute tables use the same structural pattern. A sequence tag has its own row. Child tags reference that sequence row through PARENT_IMAGING_STUDY_ATTRIBUTE_ID or PARENT_IMAGING_SERIES_ATTRIBUTE_ID. PARENT_DICOM_TAG_PATH in the study-attribute definition and PARENT_DICOM_TAG_SOURCE_VALUE in the series-attribute definition record the human-readable path to the parent sequence, including zero-based sequence indexes. PARENT_SEQUENCE_INDEX identifies the item within the immediate parent sequence.

Arrays are distinct from sequences. IS_ARRAY = 1 indicates that a tag contains multiple scalar values, and ARRAY_INDEX records the zero-based position of each value. TAG_OCCURRENCE_NUMBER distinguishes repeated occurrences of the same tag under one parent; DICOM-compliant data normally uses the default value 1.

Values are stored in the column matching their type:

Tables and Attributes

The implemented table definitions are listed below. IMAGING_STUDY and IMAGING_SERIES are available in both schemas. The two attribute tables are implemented only in CDMRADIOLOGY. Required status is the same in both variants for retained columns.

IMAGING_STUDY

One record per imaging study, linked to the patient.

Identifiable model: 35 columns. De-identified model: 15 columns.

IMAGING_SERIES

One record per DICOM series within a study, with modality, anatomy, acquisition, equipment, and aggregate image information.

Identifiable model: 44 columns. De-identified model: 35 columns.

IMAGING_STUDY_ATTRIBUTE

Study-level DICOM attributes that are not represented as dedicated columns in IMAGING_STUDY.

Identifiable model: 29 columns. De-identified model: 0 columns.

IMAGING_SERIES_ATTRIBUTE

Series-level DICOM attributes that are not represented as dedicated columns in IMAGING_SERIES.

Identifiable model: 29 columns. De-identified model: 0 columns.

Limitations

• Linkage from IMAGING_STUDY.PROCEDURE_OCCURRENCE_ID to the corresponding OMOP procedure occurrence is not currently implemented. Do not use this field to join imaging studies to PROCEDURE_OCCURRENCE until the linkage is released and validated.

Radiology Notes (De-ID)

The Windreich Department of AI and Human Health, together with Scientific Computing and Data and the BioMedical Engineering and Imaging Institute (BMEII), is pleased to announce the release of approximately 27 million de-identified radiology notes in AI Ready Mount Sinai (AIR·MS), further expanding the platform’s multimodal clinical data resources. This new dataset provides researchers with access to de-identified radiology reports that capture rich clinical interpretations of imaging studies. 

To protect patient privacy, a robust AI-assisted de-identification pipeline combining natural language processing (NLP) and large language model (LLM) technologies was developed to remove protected health information while preserving the clinical and research value of the reports.

To access the data, follow our “Getting Started” guide here: https://labs.icahn.mssm.edu/minervalab/air%E2%80%A7ms-getting-started/ 

 1. About the Dataset 

This dataset is one of the largest de-identified free-text datasets in the United States: by leveraging the extensive computing resources of Minerva together with the most current LLMs, we were able to process millions of radiology notes over the course of ~ 12 weeks. Data is stored in AIR·MS, together with other radiology data, and can be accessed using the CDMDEID.NOTE_IMAGING_XTN table in AIR·MS. The table schema can be described as follows:  

2. Compliance-Related Considerations

 2.1 IRB (Institutional Review Board) Approval – Not Required 

Since these notes have been de-identified using a number of computational approaches, IRB approval is not required to access or use this data. This means you can quickly and easily begin incorporating this data into your investigations. 

 2.2 Terms of Use/Disclaimer 

Please note that all disclaimers and terms of use that apply to AIR·MS data apply here. Additionally, ensuring the safety of de-identified data is a shared responsibility between the de-identification team and end users. De-identification of clinical notes is still a topic of active research, and methods that are 100% accurate still do not exist in peer-reviewed literature. While we have done our best to ensure our methods are best-in-class and that no PHI has escaped the masking process, some information may still be present in the de-identified data. If you do find PHI in the de-identified data, please contact us as soon as possible by submitting a ticket, so we can investigate and modify/correct/update our pipelines and tools to account for possible masking errors and issues. Additionally, data usage is subject to the AIR·MS Data Use Agreement.

3. Radiology Notes De-Identification Process 

The de-identification of radiology notes was completed through a multi-stage workflow combining traditional NLP techniques with LLMs to maximize the detection and removal of protected health information (PHI). The de-identification process included using different NLP algorithms to do an initial pass de-identifying information, and then a LLM to do a final pass. Once the note was de-identified, it was then validated using a combination of NLP and LLM methods.  

Please note: some information has been over-redacted (such as key clinical findings, medical eponyms, other results). This is a commonly occurring artifact with many de-identification pipelines (which are highly sensitive but less specific), and we will be updating the pipeline to un-redact this information in subsequent versions.  

More details and links to pre-print manuscripts will appear here shortly, so please check back soon! 

4. Human Validation and Quality Assurance

As a final quality assurance step, a representative sample of 3,000 radiology notes underwent manual review by clinical experts to verify the accuracy and completeness of the de-identification process. Overall, this AI-assisted de-identification pipeline enables secure, compliant use of radiology data for secondary research, AI model development, and multimodal analyses. 

Radiology Notes (PHI)

The Windreich Department of AI and Human Health, together with Scientific Computing and Data and the BioMedical Engineering and Imaging Institute, are pleased to announce the release of over 11 million Protected Health Information (PHI) radiology notes from 1.95 million individuals on the AIR·MS (AI Ready Mount Sinai) platform, further expanding the platform’s multimodal clinical data resources. 

To access the data, follow our “Getting Started” guide here: https://labs.icahn.mssm.edu/minervalab/air%E2%80%A7ms-getting-started/ 

This new dataset provides researchers with access to radiology reports that capture rich clinical interpretations of imaging studies. 

To protect patient privacy, IRB approval and a signed data use agreement are required to access and use this data. This enables secure, compliant use of radiology data for secondary research, AI model development, and multimodal analyses. 

Radiology Metadata (PHI)

​
AIR·MS now features radiology metadata extracted from the Mount Sinai IRW 2.0 XNAT system (via an MSDW data pipeline). This data set is comprised of detailed DICOM (Digital Imaging and Communications in Medicine) tags associated with the medical images. These tags provide essential metadata, including patient information, imaging parameters, equipment details, and procedural context, ensuring a comprehensive understanding of each radiological study. By integrating this metadata, we enable researchers to gain deeper insights into the imaging data, facilitating advanced analyses and fostering innovations in medical imaging research.

To access the data, follow our “Getting Started” guide here: https://labs.icahn.mssm.edu/minervalab/air%E2%80%A7ms-getting-started/ 

Schema: CDMRADIOLOGY
Data snapshot from 8/16/2024
Unique patients: 2,057,482

The following tables and attributes are available in AIR·MS:

Number of records: 66,993,826

Number of records: 7,745,407,087

Electroencephalography Metadata (PHI)

The CDMEEG schema contains identifiable electroencephalography (EEG) study metadata, EDF files, channel metadata, annotations, and EEG report notes. The study/file portion of the model follows a study-to-file hierarchy. Reports are loaded separately and may be associated with studies through source order identifiers when those identifiers agree.

This schema contains protected health information (PHI), including patient identifiers, names, birth dates, clinician names, clinical text, report text, and storage locations. Use it only in approved AIR·MS environments and do not export row-level data to an unapproved location.

To access the data, follow the AIR·MS “Getting Started” guide.

Schema Summary

HANA schema: CDMEEG
Tables: 5 column-store tables

Record Counts

Study start dates range from May 24, 2000 through May 26, 2026. File recording start dates range from August 29, 2005 through May 2, 2026. Report service dates range from January 1, 2024 through August 31, 2026.

Model Structure

EEG_STUDY
└── EEG_FILES
    ├── EEG_SIGNAL_METADATA
    └── EEG_ANNOTATIONS

EEG_REPORT
└── separate clinical-note feed; ORDER_ID is a possible, non-enforced study link

Of the 34,385 report rows, 20,767 have ORDER_ID populated and 16,380 match at least one EEG_STUDY.ORDER_ID. No declared foreign key or unique constraint makes this a guaranteed one-to-one link.

Keys

Tables and Columns

Required reflects HANA nullability, not whether the source value is clinically expected. HANA TIMESTAMP values do not carry a time-zone offset.

EEG_STUDY

One row per EEG study known to the source inventory. This table combines patient and encounter identifiers, acquisition metadata, clinical or review fields, source availability flags, and pipeline status.

EEG_FILES

One row per EDF file or EDF chunk associated with an EEG study. All 359,067 deployed file names have an .edf suffix.

EEG_SIGNAL_METADATA

One row per signal/channel definition in an EDF file. These fields correspond to common EDF per-signal header attributes.

EEG_ANNOTATIONS

One row per annotation extracted from an EDF file. Annotations are positioned relative to the recording timeline.

EEG_REPORT

One row per loaded EEG-related clinical note. This table is keyed by an Epic clinical-note identifier and includes OMOP linkage fields when available. It is not structurally attached to the study/file hierarchy by a declared foreign key.

Echocardiography Metadata (PHI)

AIR·MS contains DICOM metadata tags for cardiovascular imaging studies performed in the Mount Sinai Health System that are contained within the Softlink cardiovascular PACS system. This does not include radiology data contained within the radiology PACS systems. The following common modalities include US, CT, XA, NM, MR, IVUS. These modalities include, among others, echocardiographic ultrasound, vascular ultrasound, and angiographic data. These tags are linked to DICOM files by ECHO_METADATA.IMAGE_FILE_PATH in a repository on the Minerva cluster. Note that the schema name CDMECHO is a misnomer – this catalog contains much more than echocardiogram data.

To access the data, follow our “Getting Started” guide here: https://labs.icahn.mssm.edu/minervalab/air%E2%80%A7ms-getting-started/ 

The catalog contains every public DICOM metadata tag (dictionaries are publicly available, see https://dicom.innolitics.com/ciods for instance). The Tag is under ECHO_TAGS_DATA.TAGS and is labeled by its Name not its hexadecimal number for ease of readability (see https://www.dicomlibrary.com/dicom/dicom-tags/ for dictionary of public tags and name). The value of the tag is under ECHO_TAGS_DATA.TAG_VALUE.

Notes:

• As of late August 2025, there are known significant gaps in data availability around year 2019 and starting in mid 2023 and onward. The catalog ends in late 2023. Mechanisms for capturing missing data and updating the catalog moving forward are underway.
• There will be some duplicated data within the archive (i.e. 2 identical studies may be in 2 separate paths)
• There is a known bug in the way the value representation of “Person Name” is stored. It is stored as a list of single characters rather than a string (i.e. the name Smith is stored as [S,m,i,t,h])

Schema: CDMECHO

Data snapshot from: 10/17/2023

Unique patients: 885,957

Number of records: 268,682,931

Number of records: 21,089,507,788

Endoscopy Reports (PHI)

AIR·MS contains a catalog of gastroenterology endoscopy reports stored in PDF format on Minerva. The reports are for a subset of these procedures wherein the reporting system (Provation MD, Endoworks) transmits the reports as PDF files to Epic. Provation Apex and gGastro reports, including those from the Ansonia and ECNY sites are not currently included.

To access the data, follow our “Getting Started” guide here: https://labs.icahn.mssm.edu/minervalab/air%E2%80%A7ms-getting-started/ 

These are intended to be linked to the OMOP CDM via the latter’s person table. We do not currently link it to procedure records in procedure_occurrence. Various date fields related to the document and its transmission to the Epic system are available, but their validity can’t be guaranteed and filtering on those columns should be done with caution.

Number of records: 428,430

* These columns are loaded from Epic, but their accuracy may vary based on a
variety of circumstances and cannot be guaranteed

Synthetic Public Use File (DE-SynPUF)

The SYNPUF (Synthetic Public Use Files) dataset, provided by the Centers for Medicare & Medicaid Services (CMS), offers a synthetic version of Medicare claims data from the years 2008 to 2010. This dataset is meticulously designed to maintain the statistical properties and relationships present in the original data while ensuring that no actual patient information is disclosed, thereby safeguarding privacy. SYNPUF includes a comprehensive array of variables such as beneficiary demographics, chronic conditions, hospital and outpatient claims, and prescription drug events, making it an invaluable resource for researchers and data scientists. It serves as an exemplary tool for developing and testing healthcare models, algorithms, and applications without the constraints associated with sensitive real-world data. The SYNPUF dataset in AIR·MS utilizes the OMOP Common Data Model, aligned with other clinical data sets available on the platform. Since SYNPUF data does not require an approved IRB, you can easily get onboarded and start building ML models!

To access the data, follow our “Getting Started” guide here: https://labs.icahn.mssm.edu/minervalab/air%E2%80%A7ms-getting-started/ 

Schema: CDMSYNPUF
Number of patients: 2,326,856
Number of observations: 37,531,051
Number of measurements: 72,387,791