Skip to content
View on GitHub

More Datasets

A collection of classic, tabular machine-learning health datasets — the kind of clean, well-labeled data that's useful for teaching core ML concepts (classification, feature importance, class imbalance, regression) before working with the messier text and unstructured clinical data used elsewhere on this site. Each one lives in its own subfolder.

Repo: UF-AIBHS/aibhs_datasets — clone it directly, or use the per-dataset links below to download just one folder. · Download entire repo (ZIP)

Table of Contents

Overview Table

Folder Official Dataset Name Source License / Data Use Policy Notes Description
heart_disease_cleveland/ Heart Disease (Cleveland subset) UCI ML Repository CC BY 4.0 — reuse and adaptation permitted with attribution Full dataset (303 rows × 14 cols) Clinical/demographic attributes used to predict presence of heart disease; used across nearly every weekly notebook
pima_diabetes/ Pima Indians Diabetes Database jbrownlee/Datasets GitHub mirror; no longer listed in UCI's current repository No license file on the mirror repo. Originates from NIDDK (a US federal agency); cite the original NIDDK study and this mirror when reusing Full dataset (768 × 9) Diagnostic measurements from female Pima Indian patients predicting diabetes onset
heart_failure_clinical/ Heart Failure Clinical Records Dataset UCI ML Repository CC BY 4.0 — reuse and adaptation permitted with attribution Full dataset (299 × 13) Follow-up records of heart failure patients predicting mortality
breast_cancer_wisconsin/ Breast Cancer Wisconsin (Diagnostic) UCI ML Repository CC BY 4.0 — reuse and adaptation permitted with attribution Full dataset (569 × 32) Cell-nucleus features from biopsy imaging predicting malignant vs. benign
snows_cholera_map/ John Snow's 1854 Cholera Map Data Original repo (rtwilson/SnowsCholeraMap) is dead; downloaded from GitHub Gist mirrors: deaths, pumps Public domain — 1854 source data, far outside any copyright term; the Gist mirrors carry no separate license 250 addresses / 489 deaths, 8 pumps Deaths and water-pump locations from the Broad Street cholera outbreak
framingham_heart_study/ Framingham Heart Study (teaching extract) GitHub mirror of a commonly used Kaggle teaching dataset No license file on the mirror repo. The official Framingham dataset is controlled by NHLBI/BioLINCC and requires a formal data-access application for research use — this unofficial teaching extract is for coursework only, not for publication or redistribution as "the" Framingham dataset Full extract (4,240 × 16) Cardiovascular risk factors predicting 10-year coronary heart disease
nhanes/ NHANES 2017–2018, Demographics component (DEMO_J) CDC/NCHS NHANES, specifically the DEMO_J file page US federal government public-use data file — public domain domestically (17 U.S.C. §105); free to use, no permission or fee required Only 1 of many component files for 1 of many survey cycles (9,254 × 46); converted from .XPT to .csv National survey demographics + sampling weights; not yet joined to any health/exam/lab component
brfss/ Behavioral Risk Factor Surveillance System, 2023 CDC BRFSS US federal government public-use data file — public domain domestically (17 U.S.C. §105); free to use, no permission or fee required Subsampled: random 5,000 of 433,323 rows (random_state=42); full file is ~1.2 GB unzipped Annual US health-behavior survey (350 variables)
life_expectancy_who/ Life Expectancy (WHO) GitHub Gist mirror; exact upstream Kaggle version unclear (see below) Unclear — the two candidate Kaggle originals carry different licenses (one "Other/unspecified," one CC0 Public Domain) and this file doesn't cleanly match either; treat as coursework-only until provenance is confirmed 2,461 × 23, 156 countries; doesn't match either well-known Kaggle release of this dataset Country-year panel (2000–2015) of health/economic indicators predicting life expectancy
diabetes_130_hospitals/ Diabetes 130-US Hospitals for Years 1999–2008 UCI ML Repository CC BY 4.0 — reuse and adaptation permitted with attribution Full dataset (101,766 × 50) + IDS_mapping.csv lookup table De-identified inpatient encounters predicting 30-day readmission
stroke_prediction/ Stroke Prediction Dataset Kaggle (via GitHub Gist mirror) Kaggle-listed as "Data files © Original Authors" — not CC0; no redistribution rights granted beyond Kaggle's terms, use for coursework with attribution to the Kaggle source Full dataset (5,110 × 12) Demographic/lifestyle attributes predicting stroke occurrence (~5% positive class)
mimic_iv_demo/ MIMIC-IV Clinical Database Demo v2.2 PhysioNet ODC Open Database License v1.0 (ODbL) — open access, no PhysioNet credentialing or CITI training required (unlike full MIMIC-IV) Full demo (100 patients, 26 tables, ~16 MB zipped); excludes free-text clinical notes De-identified EHR extract from Beth Israel Deaconess Medical Center; lets you prototype against MIMIC-IV's schema before requesting full credentialed access
eicu_demo/ eICU Collaborative Research Database Demo v2.0.1 PhysioNet ODC Open Database License v1.0 (ODbL) — open access, no PhysioNet credentialing or CITI training required Full demo (~137 MB uncompressed): 31 gzip-compressed CSVs plus a SQLite build (sqlite/eicu_v2_0_1.sqlite3.gz, 78 MB) De-identified ICU stays (2,500+) from 20+ US hospitals: vitals, labs, care-plan and APACHE severity data

A Note on Licensing

The License / Data Use Policy column above reflects what each source publisher states as of this writing. A few things worth knowing:

  • This is not legal advice. When a folder's status is "unclear" or "no license file," don't redistribute that data outside coursework or claim it as your own — cite the original source and, where one exists, follow that source's own terms.
  • "Public domain" for CDC/NCHS data (NHANES, BRFSS) means the data itself carries no copyright restriction in the US; it does not waive CDC's usual data-quality and citation guidance.
  • Datasets sourced from PhysioNet's demo projects (MIMIC-IV, eICU) are open access under ODbL, but the full, non-demo versions of those databases require PhysioNet credentialing and CITI "Data or Specimens Only Research" training — don't assume demo-level access extends to the full database.
  • If a license changes upstream, this table may lag — check the linked source page for the current, authoritative terms before using a dataset outside this project.

Subsampled & Partial Datasets

Everything else in this folder is the complete dataset as published. Two are not, for different reasons: one was cut down to a random sample of rows, the other only has a slice of a much bigger program pulled in. Both are called out explicitly below so nobody mistakes either for the complete dataset.

Folder Original size Size present here What's missing and why
brfss/ 433,323 rows × 350 columns (~1.2 GB unzipped) 5,000 rows × 350 columns, 5.6 MB (LLCP2023_sample5000.csv) A true random subsample: df.sample(n=5000, random_state=42) on the full 2023 file. The 1.2 GB original was too big to keep in the repo, so it was downloaded, sampled, and deleted rather than committed. Re-running the same seed against the CDC's 2023 file reproduces this exact sample.
nhanes/ The full NHANES program: 12+ two-year survey cycles since 1999, each split into 5 broad components (Demographics, Dietary, Examination, Laboratory, Questionnaire) that in turn span dozens of individual files per cycle (the Laboratory component alone covers 800+ distinct lab tests) 1 cycle (2017–2018), 1 file (Demographics, DEMO_J): 9,254 rows × 46 columns This isn't a row-level subsample, it's the complete demographics file for one cycle. No exam, lab, diet, or questionnaire file has been pulled for any cycle, and no other year has been downloaded. Everything present is real and unmodified; it's just a narrow slice of what NHANES actually covers.

1. Heart Disease (Cleveland)

File: processed.cleveland.data, 303 rows by 14 columns, no header row. · Download this folder

Demographic columns: age, sex

Health columns: cp (chest pain type), trestbps (resting blood pressure), chol (serum cholesterol), fbs (fasting blood sugar over 120 mg/dl), restecg (resting ECG results), thalach (max heart rate achieved), exang (exercise-induced angina), oldpeak (ST depression from exercise), slope (slope of the peak exercise ST segment), ca (number of major vessels seen on fluoroscopy), thal (thalassemia test result), and target (disease status, 0 to 4; commonly collapsed to 0/1 for binary classification).

Robert Detrano and collaborators put this together in 1988 across four sites: Cleveland Clinic Foundation, the Hungarian Institute of Cardiology, University Hospital Zurich/Basel, and the V.A. Medical Center in Long Beach. The question was whether ordinary, non-invasive clinical measurements could stand in for an angiogram when flagging coronary artery disease. Cleveland's data has the fewest gaps of the four, which is why it's the one almost everyone actually uses.

2. Pima Indians Diabetes

File: pima-indians-diabetes.data.csv, 768 rows by 9 columns, no header row. · Download this folder

Demographic columns: every subject is a female of Pima Indian heritage; age is recorded per row.

Health columns: pregnancies, glucose (2-hour oral glucose tolerance test), blood_pressure (diastolic, mm Hg), skin_thickness (triceps skinfold, mm), insulin (2-hour serum insulin), bmi, diabetes_pedigree (a score for diabetes likelihood based on family history), and outcome (0/1).

The data comes from the NIDDK's long-running study of the Pima community near Phoenix, Arizona, a population that's been tracked since the 1960s because it has one of the highest recorded rates of type 2 diabetes anywhere. Smith and colleagues used it in 1988 to test whether routine clinical measurements could forecast diabetes onset, and it's been a standard classroom dataset ever since.

3. Heart Failure Clinical Records

File: heart_failure_clinical_records_dataset.csv, 299 rows by 13 columns. · Download this folder

Demographic columns: age, sex

Health columns: anaemia, creatinine_phosphokinase (a muscle-damage enzyme), diabetes, ejection_fraction (percent of blood the heart pumps out per beat), high_blood_pressure, platelets, serum_creatinine, serum_sodium, smoking, time (days of follow-up), and DEATH_EVENT.

Collected at the Faisalabad Institute of Cardiology and the Allied Hospital in Faisalabad, Pakistan, between April and December 2015, from patients who already had heart failure with reduced ejection fraction. Ahmad and colleagues published it first; Chicco and Jurman revisited it in 2020 (BMC Medical Informatics and Decision Making) and made the case that two numbers, serum creatinine and ejection fraction, do most of the work in predicting whether a patient survives. That claim is probably why the dataset shows up so often in feature-importance exercises.

4. Breast Cancer Wisconsin (Diagnostic)

File: wdbc.data, 569 rows by 32 columns, no header row. · Download this folder

Demographic columns: none besides a patient id.

Health columns: diagnosis (M or B) plus 30 numeric features: 10 measurements (radius, texture, perimeter, area, smoothness, compactness, concavity, concave_points, symmetry, fractal_dim), each reported three ways (mean, standard error, and worst value) across the cell nuclei visible in one image.

Dr. William Wolberg at the University of Wisconsin Hospitals collected fine needle aspirate biopsies of breast masses; Street, Wolberg, and Mangasarian built the "Xcyt" program that turned each digitized slide into these shape and texture numbers. The goal was a diagnosis that didn't require surgery. It's held up well: this is still a reliable small classification dataset, clean enough that model performance rarely comes down to data quality rather than modeling choices.

5. Snow's Cholera Map

Files: deaths.csv (250 addresses), pumps.csv (8 pumps). · Download this folder

This isn't a per-patient dataset. Each row in deaths.csv is a street address with a death count (Number of deaths) and a location (X coordinate, Y coordinate); the 250 addresses here sum to 489 deaths. pumps.csv gives pump locations, but the Pump Name field is unpopulated in this mirror (every row reads -999), so no actual pump names survived into this copy.

John Snow mapped these deaths during the 1854 cholera outbreak around Broad Street in Soho, London, where several hundred people died within about ten days. Plotting deaths by address showed them clustered tightly around one pump, the argument Snow used against the era's dominant "miasma" theory, which blamed bad air rather than contaminated water. It's a genuinely foundational moment for epidemiology and for the broader idea of mapping data to trace a cause.

6. Framingham Heart Study

File: framingham.csv, 4,240 rows by 16 columns. · Download this folder

Demographic columns: male, age, education

Health columns: currentSmoker, cigsPerDay, BPMeds, prevalentStroke, prevalentHyp, diabetes, totChol, sysBP, diaBP, BMI, heartRate, glucose, and TenYearCHD (the target).

Started in 1948 in Framingham, Massachusetts, with 5,209 adults who had no prior history of heart disease or stroke. It's the longest-running cardiovascular cohort study going, later expanded with an Offspring Cohort in 1971 and a Third Generation Cohort in 2002, and it's the study that gave medicine the term "risk factor" in the first place, after investigators showed smoking, cholesterol, blood pressure, and obesity all raised cardiovascular risk. This CSV is a commonly used teaching extract with a single 10-year CHD outcome, which works for both correlation exercises and classification.

7. NHANES (2017–2018 Demographics)

Files: DEMO_J.xpt (original SAS transport format), DEMO_J_2017-2018.csv (converted). 9,254 rows by 46 columns. · Download this folder

Demographic columns: RIAGENDR (gender), RIDAGEYR (age), RIDRETH1/RIDRETH3 (race/ethnicity), DMDEDUC2/DMDEDUC3 (education), DMDMARTL (marital status), DMDHHSIZ (household size), INDHHIN2/INDFMIN2 (income bracket), INDFMPIR (income-to-poverty ratio).

The rest of the columns are survey machinery: SDDSRVYR (cycle), RIDSTATR (interview/exam status), WTINT2YR/WTMEC2YR (sample weights), SDMVPSU/SDMVSTRA (variance strata). NHANES uses a complex multistage sample, not a simple random one, so these weight and strata columns are required for any population-level estimate rather than optional extras.

This file only covers demographics and weights for one two-year cycle. The actual exam results, labs, and questionnaires live in separate component files for the same cycle, joined on SEQN.

NHANES traces back to the National Health Examination Survey of the early 1960s and has run as a continuous annual survey since 1999, combining household interviews with an in-person exam and lab work on roughly 5,000 people a year. It's the source behind U.S. growth charts and national diabetes and obesity estimates, and its explicit sampling weights make it a good fit for sampling and confidence-interval exercises.

8. BRFSS (2023, subsampled)

File: LLCP2023_sample5000.csv, 5,000 rows by 350 columns, a random sample (random_state=42) of the full 433,323-row file. · Download this folder

Demographic columns: _STATE, SEXVAR, MARITAL, EDUCA, _AGE80, _RACE, INCOME2

Health columns (a few of roughly 340): GENHLTH (self-rated general health), PHYSHLTH/MENTHLTH (poor-health days in the past month), EXERANY2 (any exercise), CVDINFR4/CVDCRHD4 (heart attack / coronary heart disease history), DIABETE4, SMOKE100, _RFHYPE6 (high blood pressure), _BMI5CAT. The rest cover chronic conditions, screenings, immunizations, disability, and healthcare access.

BRFSS has been run by the CDC since 1984 and is the largest continuously conducted health survey in the world: an annual phone survey administered state by state to track health behaviors and chronic conditions among US adults, feeding directly into public-health funding decisions. The full 2023 file is about 1.2 GB unzipped with over 400,000 respondents, more than most coursework needs, so what's here is a fixed random sample of 5,000 rows.

9. WHO Life Expectancy

File: Life_Expectancy_Data.csv, 2,461 rows by 23 columns, covering 156 countries from 2000 to 2015. · Download this folder

Demographic columns: Country, Continent, Year, Status (Developing/Developed), Population, Income_composition_of_resources, Schooling

Health columns: Life_expectancy (the target), Adult_Mortality, infant_deaths, under_five_deaths, Alcohol, BMI, Hepatitis_B/Measles/Polio/Diphtheria (immunization coverage), HIV/AIDS, thinness 1-19 years/thinness 5-9 years, percentage_expenditure/Total_expenditure, GDP.

A provenance note: this file doesn't cleanly match either of the two well-known Kaggle releases of "Life Expectancy (WHO)." The original (kumarajarshi) has 2,938 rows and 22 columns with no Continent field; a later cleaned version (lashagoch) has 2,864 rows across 179 countries. This file lands at 2,461 rows across 156 countries and does include Continent, so it's most likely a further-filtered derivative of one of those two rather than an exact copy of either. Worth knowing before citing it as "the WHO dataset" in a project write-up.

It's a secondary compilation drawing on the WHO Global Health Observatory and UN economic data, not an official WHO release, built to study which health, immunization, and economic factors track with a country's life expectancy. What makes it worth including here is that its target is continuous, unlike most of the other datasets in this collection — a good fit for regression exercises rather than classification.

10. Diabetes 130-US Hospitals (1999–2008)

Files: diabetic_data.csv (101,766 rows by 50 columns), IDS_mapping.csv (lookup tables decoding admission_type_id, discharge_disposition_id, and admission_source_id). · Download this folder

Demographic columns: race, gender, age (10-year bands), weight (mostly unrecorded)

Health columns: admission_type_id, discharge_disposition_id, admission_source_id, time_in_hospital, num_lab_procedures, num_procedures, num_medications, number_outpatient/number_emergency/number_inpatient (prior visit counts), diag_1/diag_2/diag_3 (ICD-9 codes), number_diagnoses, max_glu_serum, A1Cresult, 23 individual medication columns (metformin, insulin, glipizide, and so on), change, diabetesMed, and readmitted (the target: no, under 30 days, or over 30 days).

Pulled from the Cerner Health Facts database: de-identified inpatient records from 130 US hospitals and delivery networks, 1999 through 2008, limited to encounters involving a diabetes diagnosis. Strack and colleagues put it together in 2014 (BioMed Research International) to check whether measuring HbA1c during a hospital stay lowers the odds of an early readmission. It's a lot bigger and a lot messier than most of the other datasets here: missing weight data, ICD-9 codes instead of readable labels, two dozen near-identical medication columns. That makes it a good "real data is messy" exercise.

11. Stroke Prediction Dataset

File: healthcare-dataset-stroke-data.csv, 5,110 rows by 12 columns. · Download this folder

Demographic columns: gender, age, ever_married, work_type, Residence_type

Health columns: hypertension, heart_disease, avg_glucose_level, bmi, smoking_status, stroke (the target, around 5% positive).

This one is published on Kaggle under the name fedesoriano, and unlike everything else in this collection, there's no clear record of where the data actually came from or how it was collected. Worth knowing before using it for a project — not knowing the collection method makes it harder to say what the results actually generalize to. It earns its spot here mostly for the class imbalance: only about 5% of cases are stroke-positive, which makes it a useful counterexample to the more balanced datasets elsewhere in this collection, good for teaching precision, recall, and ROC-AUC instead of leaning on plain accuracy.

12. MIMIC-IV Clinical Database Demo

Folder: mimic_iv_demo/, 26 tables across hosp/ (hospital-wide EHR: admissions, diagnoses, labs, prescriptions, microbiology, etc.) and icu/ (ICU-specific: chart events, inputs/outputs, procedures), all gzip-compressed CSVs. ~16 MB total. · Download this folder

License: ODC Open Database License v1.0 — open access. Unlike the full MIMIC-IV database, the demo requires no PhysioNet credentialing or CITI training; anyone can use the files under the license terms.

The Medical Information Mart for Intensive Care (MIMIC)-IV database contains deidentified electronic health records from patients admitted to Beth Israel Deaconess Medical Center in Boston. Access to the full database is restricted to credentialed PhysioNet users who complete human-subjects research training, because it covers hundreds of thousands of real admissions. This demo is an openly-available 100-patient subset with the same 26-table schema as the full database, minus free-text clinical notes, published specifically so instructors and researchers can prototype against MIMIC-IV's structure or run a workshop without going through credentialing first. If a project outgrows the demo, the natural next step is a PhysioNet account, CITI training, and a formal data-use request for the full database at physionet.org/content/mimiciv.

13. eICU Collaborative Research Database Demo

Folder: eicu_demo/, 31 gzip-compressed CSVs plus a SQLite build of the same data (sqlite/eicu_v2_0_1.sqlite3.gz). ~137 MB uncompressed — well past every other dataset here (the next largest, Diabetes 130-US Hospitals, is 18 MB), but every individual file is under GitHub's 100 MB per-file limit, so it's included directly. · Download this folder

License: ODC Open Database License v1.0 — open access, same as the MIMIC-IV demo above: no PhysioNet credentialing or CITI training required. Authoritative source: physionet.org/content/eicu-crd-demo/2.0.1.

The eICU Collaborative Research Database pools ICU data contributed by Philips Healthcare and the MIT Laboratory for Computational Physiology from over 200 US hospitals using the Philips eICU telehealth program. This demo covers 2,500+ ICU stays from 20+ of the larger contributing hospitals: vital signs, care-plan documentation, APACHE severity-of-illness scores, diagnoses, lab results, medications, and nursing assessments, all deidentified to HIPAA Safe Harbor standards. As with MIMIC-IV, the full (non-demo) eICU database requires PhysioNet credentialing; the demo does not, which is what makes it usable without a data-access request.