MIMIC-III-Ext-SBDH: An annotated social and behavioral determinants of health dataset

@article{PhysioNet-mimic-iii-ext-sbdh-1.0.0,
  author = {Gu, Zifan and He, Lesi and Yang, Donghan},
  title = {{MIMIC-III-Ext-SBDH: An annotated social and behavioral determinants of health dataset}},
  journal = {{PhysioNet}},
  year = {2026},
  month = sep,
  note = {Version 1.0.0},
  doi = {10.13026/ej3k-2436},
  url = {https://doi.org/10.13026/ej3k-2436}
}

Abstract

This study presents MIMIC-III-Ext-SBDH, a manually reviewed and consolidated annotation dataset of social and behavioral determinants of health (SBDH) derived from clinical notes in the MIMIC-III database. We previously developed SBDH-Reader, a large language model (LLM)-powered method, to extract granular SBDH data from clinical notes through prompt engineering. During the iterative process of refining the SBDH-Reader prompt, we identified several instances where the original ground truth labels, derived from two previously annotated datasets, were inaccurate. These inconsistencies were systematically reviewed and corrected through consensus-based manual annotation, resulting in a consolidated and updated corpus. The final dataset includes 7,225 clinical notes from 6,382 patients across six SBDH categories: employment, housing, marital status, and substance use including alcohol, tobacco, and drug use. The release includes note-level annotations, associated metadata, and checksum files for data integrity verification.

MIMIC-III-Ext-SBDH is intended to support the development and evaluation of natural language processing and (agentic) LLM methods for extracting SBDH information from unstructured clinical narratives and to facilitate reproducible research on socially informed clinical prediction and phenotyping.

Background

Social and behavioral determinants of health (SBDH) encompass a broad range of factors that describe an individual’s living conditions at the personal, community, and societal levels. A growing body of research highlights the significance of integrating SBDH to enhance clinical risk assessment, outcome prediction, and clinical trial enrollment, ultimately guiding therapeutic interventions across various clinical settings [1-8].

Despite the recognized significance of SBDH in the medical domain, access to these patient-level characteristics is limited due to the lack of effective methods for high-quality and continuous data collection, particularly in clinical settings [9,10]. Previous studies have attempted to collect these data through patient surveys delivered during routine care, but the scalability and generalizability of this approach is limited [1,12]. Many electronic health records (EHRs) provide structured data fields for SBDH, but these require providers to enter and are often poorly populated [9,10,13]. Information about SBDH is often documented in the narrative sections of clinical notes, but extracting detailed SBDH information from unstructured notes has proven challenging [1,9,10,14-17].

Our goal for sharing this dataset is to provide other credentialed researchers with access to a manually validated corpus of structured SBDH annotations from clinical notes. This dataset combines two previously annotated corpora [18,19], each of which has been systematically reviewed and corrected to address label inaccuracy. In our accompanying publication, we demonstrate how this resource can be used in prompt engineering-based strategies, performing both multi-label (i.e., across multiple SBDH categories) and multi-class classification (i.e., across multiple possible attributes) in a single prompt. This offers a more flexible, versatile, and generalizable solution for extracting SBDH data from real-world clinical notes.

Methods

The ground truth review and correction process was conducted in parallel with the iterative prompt refinement for SBDH-Reader. We first describe the prompt refinement process, followed by the procedure used to correct ground truth labels.

To develop and refine the SBDH-Reader prompt using two previously annotated MIMIC-III datasets, we iteratively added and modified specifications for each SBDH category and attribute present in these two datasets. For each low-performing attribute, we compared the original notes with the supporting evidence text extracted by SBDH-Reader. If misclassifications made by SBDH-Reader were found to be repeatable across multiple notes, indicating systematic rather than isolated errors, we incorporated new specifications into the prompt to prevent such errors. The iterative refinement continued until minimal improvement was observed from one iteration to the next, typically when remaining errors were case-specific and not generalizable.

During this manual error analysis, we identified several instances where SBDH-Reader correctly classified an attribute with valid supporting evidence, but the original ground truth labels in the two previously annotated MIMIC-III datasets were inaccurate. In these cases, a panel of three data scientists reviewed the notes and resolved annotation discrepancies through consensus to determine the final updated label. We then documented the note ID, the original label, and the corrected label in a tracking table. For this release, we provide the note ID and note text from MIMIC-III and the consolidated, updated annotation labels, formatted consistently with the corresponding original dataset.

Dataset preprocessing, aggregation, and verification were performed in a Python 3.9 environment managed with Conda. Structured data processing and note-level aggregation were implemented using pandas (v2.0.3) and NumPy (v1.24.3). Compatibility testing was performed on Red Hat Enterprise Linux 7.9 and Red Hat Enterprise Linux 9.6 (Plow) systems. SHA-256 checksum files were generated to verify dataset integrity prior to release.

Data Description

The first dataset, established by Guevara et al. (MIMIC-G) [19], contains 200 notes written by physicians, nurses, and social workers for 183 patients. To generate full-length notes as input data for SBDH-Reader, we merged the original sentence-level texts (5328 entries) and ground truths from the MIMIC-G dataset into the note level (200 entries). The second dataset, established by Ahsan et al. (MIMIC-A) [18], contains 7025 discharge summaries for 6199 patients. Following the original annotation method, we only used “social history” section as input data [18].

Note that for consistency with the source datasets, we retained the original annotation formats rather than harmonizing column names across datasets. This choice minimizes transition overhead for users from the original datasets to our version.

The manual review process resulted in three rounds of ground truth revision for MIMIC-A and two rounds for MIMIC-G. In total, 37 instances in MIMIC-G and 31 instances in MIMIC-A were corrected. Below is a summary of ground truth corrections in the MIMIC-G and MIMIC-A datasets.

MIMIC3_SBDH_2023_Guevara_rev

Main updates

• A new note-level file was created by aggregating the original sentence-level annotations (5,328 records) into 200 note-level records.
• A total of 37 original annotations were modified based on the in-house review of the original notes detailed in Methods above.
• Only 3 SDOH categories were included in the note-level file: 1. Housing 2. Relationship 3. Employment.
• A sha256.txt to verify data integrity.

Metadata Variables

• provider_type: profession of the individual who authored the clinical note.
• patient_id: equivalent to SUBJECT_ID in MIMIC-III.
• note_id: equivalent to ROW_ID in MIMIC-III.
• sentence_index: order in which the sentence appears within the corresponding note_id.
• text: plain-text content of the sentence.

SDOH Annotation Variables and Allowed Values

SDOH annotation variables follow the naming convention CATEGORY_attribute, where CATEGORY represents the broader SDOH domain and attribute represents the specific observed status within that category. Each variable is represented as a binary indicator:

• 0 = absence of the category/attribute
• 1 = presence of the category/attribute

The PARENT category is binary and does not contain attribute subcategories.

SDOH Categories and Allowed Attributes

1. Employment status
   • employed
   • unemployed
   • underemployed
   • retired
   • disability
   • student
2. Housing issue
   • financial_status
   • undomiciled
   • other
3. Transportation issue
   • distance
   • resource_limitation
   • other
4. Parental status
   • PARENT (presence of a child under 18 years old)
5. Relationship status
   • married
   • partnered
   • widowed
   • divorced
   • single
6. Social support
   • plus (presence of social support)
   • minus (absence of social support)

For example, EMPLOYMENT_unemployed = 1 indicates that unemployment was identified in the corresponding clinical note. For more detailed data description, see MIMIC3_SBDH_2023_Guevara_rev/README.md.

MIMIC3_SBDH_2021_Ahsan_rev

Main updates

• Raw note text from MIMIC-III was added.
• A total of 31 original annotations were modified where necessary, based on an in-house review of the original notes.
• Text column SOCIAL_HISTORY_TEXT using regular expressions (RE) to capture the "Social History" section, defined as any content between the headers "Social History" and "Physical Exam". If RE matching fails, the entire note is pasted.
• A sha256.txt to verify data integrity.

Metadata Variables

• ROW_ID: original ROW_ID identifier from MIMIC-III.
• TEXT: full discharge summary note text.
• SOCIAL_HISTORY_TEXT: text extracted from the “Social History” section using regular expression-based parsing.

SBDH Annotation Variables and Allowed Values

·Community-Present (0: False, 1: True)

• True: There was at least one passage related to active social support
• False: No related passage was identified

· Community-Absent (0: False, 1: True)

• True: There was at least one passage related to the loss of social support
• False: No related passage was identified

· Education (0: False, 1: True)

• True: There was at least one passage related to the patient’s education such as schooling, college, or degree attainment
• False: No related passage was identified

· Economics (0: None, 1: True, 2: False)

• True: The patient was currently employed
• False: The patient was unemployed (including retirement)
• None: No related passage was identified

· Environment (0: None, 1: True, 2: False)

• True: There was at least one passage related to the presence of housing
• False: There was at least one passage related to the lack of housing
• None: No related passage was identified

· Alcohol (0: None, 1: Present, 2: Past, 3: Never, 4: Unsure)

• Present: The patient was a current alcohol user
• Past: The patient consumed alcohol in the past and had quit
• Never: The patient had never consumed alcohol
• Unsure: The discharge summary contained an ambiguous passage related to alcohol consumption
• None: No related passage was identified

· Tobacco (0: None, 1: Present, 2: Past, 3: Never, 4: Unsure)

• Definitions are identical to those used for Alcohol Use

· Drug (0: None, 1: Present, 2: Past, 3: Never, 4: Unsure)

• Definitions are identical to those used for Alcohol Use

Each ROW_ID in MIMIC-SBDH.csv corresponds to a single unique discharge summary. Although the raw file contains over two million physical lines when counted using line-based utilities (e.g. wc -l), this is due to embedded newline characters within quoted note text fields in the CSV format. When parsed using standard CSV readers such as pandas, the file contains 7,025 unique note-level records. For more detailed data description, see MIMIC3_SBDH_2021_Ahsan_rev/README.md.

Usage Notes

Here we show an example on how to load and inspect both datasets downloaded from PhysioNet.

from pathlib import Path
import pandas as pd

# Set this to the PhysioNet download base directory.
# The base directory should contain:
#   MIMIC3_SBDH_2021_Ahsan_rev/MIMIC-SBDH.csv
#   MIMIC3_SBDH_2023_Guevara_rev/SDOH_MIMICIII_physio_release_note.csv
data_dir = Path(".")

# Define file paths using the PhysioNet folder structure.
mimic_sbdh_path = data_dir / "MIMIC3_SBDH_2021_Ahsan_rev" / "MIMIC-SBDH.csv"
release_note_path = data_dir / "MIMIC3_SBDH_2023_Guevara_rev" / "SDOH_MIMICIII_physio_release_note.csv"

# Load the CSV files.
# pandas correctly handles quoted multiline text fields in the CSV files.
mimic_sbdh = pd.read_csv(mimic_sbdh_path)
release_note = pd.read_csv(release_note_path)

# Preview the first few rows of each dataset.
print("First rows of MIMIC-SBDH.csv:")
print(mimic_sbdh.head(3).to_string(index=False))
print()

print("First rows of SDOH_MIMICIII_physio_release_note.csv:")
print(release_note.head(3).to_string(index=False))
print()

This dataset is intended for use in the development and validation of both LLM-based and traditional NLP methods for extracting SBDH data from clinical notes. The manually validated SBDH annotations offer ground truths in clinical information extraction and text classification tasks. We have also shown in our accompanying publication that the SBDH-Reader developed using this dataset has the ability to quickly adapt to new SBDH terminologies in a granular and flexible manner during inference.

A key limitation of this dataset is that SBDH information documented by individual providers may be incomplete, inaccurate, or reflect implicit bias. Furthermore, the high prevalence of missing SBDH documentation in real-world clinical notes presents challenges for interpretation and model development. Additionally, this dataset represents a snapshot from a specific geographic region during a defined time window, which may limit generalizability. We encourage researchers to explore methods for integrating unstructured notes, structured EHR data, and area-level SBDH indicators to construct more comprehensive and robust patient-level SBDH profiles.

Release Notes

Version 1.0.0: Initial public release of the dataset.

Ethics

Data from MIMIC-III were de-identified at the source [20]. The LLM used in this study, GPT-4o, was operated on the University of Texas Southwestern Medical Center’s private Microsoft Azure OpenAI service. Per contractual agreements with Microsoft, data processed on this service will remain inaccessible to OpenAI or other customers and will not be used to improve OpenAI models or any Microsoft or 3rd party products or services. Use of MIMIC-III data in this setting is in compliance with the PhysioNet Credentialed Health Data Use Agreement 1.5.0 [21] and the policy of “Responsible use of MIMIC data with online services like GPT” [22]. This study protocol was approved by UTSW’s institutional review board (Protocol #STU-2024-0087).

Acknowledgements

This work was supported by the Texas Health Resources Clinical Scholars Program (DMY).

Conflicts of Interest

The authors have no competing interests to declare.

References

1. Chen M, Tan X, Padman R. Social determinants of health in electronic health records and their impact on analysis and risk prediction: A systematic review. J Am Med Inform Assoc. Nov 1 2020;27(11):1764-1773. doi:10.1093/jamia/ocaa143
2. Lu MLR, Davila CD, Shah M, et al. Marital status and living condition as predictors of mortality and readmissions among African Americans with heart failure. Int J Cardiol. Nov 1 2016;222:313-318. doi:10.1016/j.ijcard.2016.07.185
3. Sterling MR, Ringel JB, Pinheiro LC, et al. Social Determinants of Health and 90-Day Mortality After Hospitalization for Heart Failure in the REGARDS Study. J Am Heart Assoc. May 5 2020;9(9):e014836. doi:10.1161/JAHA.119.014836
4. Segar MW, Hall JL, Jhund PS, et al. Machine Learning-Based Models Incorporating Social Determinants of Health vs Traditional Models for Predicting In-Hospital Mortality in Patients With Heart Failure. JAMA Cardiol. Aug 1 2022;7(8):844-854. doi:10.1001/jamacardio.2022.1900
5. Hiatt RA, Breen N. The social determinants of cancer - A challenge for transdisciplinary science. American Journal of Preventive Medicine. Aug 2008;35(2):S141-S150. doi:10.1016/j.amepre.2008.05.006
6. Adkins-Jackson PB, George KM, Besser LM, et al. The structural and social determinants of Alzheimer's disease related dementias. Alzheimers Dement. Jul 2023;19(7):3171-3185. doi:10.1002/alz.13027
7. Sekar RR, Herrel LA, Stensland KD. Social Determinants of Health and the Availability of Cancer Clinical Trials in the United States. JAMA Netw Open. May 1 2024;7(5):e2410162. doi:10.1001/jamanetworkopen.2024.10162
8. Rae S, Shaya S, Taylor E, et al. Social determinants of health inequalities in early phase clinical trials in Northern England. Br J Cancer. Sep 2024;131(4):685-691. doi:10.1038/s41416-024-02765-w
9. Hatef E, Rouhizadeh M, Tia I, et al. Assessing the Availability of Data on Social and Behavioral Determinants in Structured and Unstructured Electronic Health Records: A Retrospective Analysis of a Multilevel Health Care System. JMIR Med Inform. Aug 2 2019;7(3):e13802. doi:10.2196/13802
10. Truong HP, Luke AA, Hammond G, Wadhera RK, Reidhead M, Joynt Maddox KE. Utilization of Social Determinants of Health ICD-10 Z-Codes Among Hospitalized Patients in the United States, 2016-2017. Med Care. Dec 2020;58(12):1037-1043. doi:10.1097/MLR.0000000000001418
11. Dupre ME, Nelson A, Lynch SM, et al. Socioeconomic, Psychosocial and Behavioral Characteristics of Patients Hospitalized With Cardiovascular Disease. Am J Med Sci. Dec 2017;354(6):565-572. doi:10.1016/j.amjms.2017.07.011
12. Dupre ME, Nelson A, Lynch SM, et al. Identifying Nonclinical Factors Associated With 30-Day Readmission in Patients with Cardiovascular Disease: Protocol for an Observational Study. JMIR Res Protoc. Jun 15 2017;6(6):e118. doi:10.2196/resprot.7434
13. Guo Y, Chen Z, Xu K, et al. International Classification of Diseases, Tenth Revision, Clinical Modification social determinants of health codes are poorly used in electronic health records. Medicine (Baltimore). Dec 24 2020;99(52):e23818. doi:10.1097/MD.0000000000023818
14. Vest JR, Grannis SJ, Haut DP, Halverson PK, Menachemi N. Using structured and unstructured data to identify patients' need for services that address the social determinants of health. Int J Med Inform. Nov 2017;107:101-106. doi:10.1016/j.ijmedinf.2017.09.008
15. Patra BG, Sharma MM, Vekaria V, et al. Extracting social determinants of health from electronic health records using natural language processing: a systematic review. J Am Med Inform Assoc. Nov 25 2021;28(12):2716-2727. doi:10.1093/jamia/ocab170
16. Mullangi S, Aviki EM, Hershman DL. Reexamining Social Determinants of Health Data Collection in the COVID-19 Era. JAMA Oncol. Dec 1 2022;8(12):1736-1738. doi:10.1001/jamaoncol.2022.4543
17. Ong JCL, Seng BJJ, Law JZF, et al. Artificial intelligence, ChatGPT, and other large language models for social determinants of health: Current state and future directions. Cell Rep Med. Jan 16 2024;5(1):101356. doi:10.1016/j.xcrm.2023.101356
18. Ahsan H, Ohnuki E, Mitra A, Yu H. MIMIC-SBDH: A Dataset for Social and Behavioral Determinants of Health. Proc Mach Learn Res. Aug 2021;149:391-413.
19. Guevara M, Chen S, Thomas S, et al. Large language models to identify social determinants of health in electronic health records. NPJ Digit Med. Jan 11 2024;7(1):6. doi:10.1038/s41746-023-00970-0
20. Johnson AE, Pollard TJ, Shen L, et al. MIMIC-III, a freely accessible critical care database. Sci Data. May 24 2016;3:160035. doi:10.1038/sdata.2016.35
21. PhysioNet. PhysioNet Credentialed Health Data License 1.5.0. Accessed 2/19/2025, https://physionet.org/about/licenses/physionet-credentialed-health-data-license-150/
22. PhysioNet. Responsible use of MIMIC data with online services like GPT. Accessed 1/19/2025. https://physionet.org/news/post/gpt-responsible-use