Datasets:
G-Health Risk Assessment Datasets
This repository contains the public source datasets used to train and evaluate the specialist prediction models of G-Health. The manuscript uses 18 task families in four task types: 12 tabular classification tasks, 2 tabular regression tasks, 3 image-classification families and 1 object-detection task.
All datasets are third-party public collections. Each retains the license and terms of its original source (linked below). The Apache-2.0 license of this repository applies to the organization and documentation added here.
Prediction labels
Each model predicts the label defined in its source dataset. The labels differ in what they mean and when they apply:
- Current: a condition present at the time of measurement (for example, a diagnosis, a biopsy result or prevalent hypertension).
- Follow-up outcome: an event after the measurements (death or recurrence).
- Calculated score: a value computed by a published formula from the included inputs.
- Lesion class: the class and location of a lesion in an image.
Class probabilities, regression outputs and detection confidence scores produced by the G-Health models are model outputs for these source labels. They were evaluated for discrimination under outer five-fold cross-validation but were not calibrated to any target population, and they should not be read as individual probabilities of disease.
Tabular classification
| ID | Task | Samples | Classes | Prediction label | Timing | Source design | Source |
|---|---|---|---|---|---|---|---|
| RISK-01 | Obesity | 2,111 | 7 | Obesity level by body-mass-index band | Current | Web survey, Mexico, Peru and Colombia | UCI 544 |
| RISK-02 | Diabetes | 253,680 | 2 | Self-reported diabetes or prediabetes | Current | US telephone health survey (CDC BRFSS 2015) | Diabetes Health Indicators |
| RISK-03 | Hepatitis (mortality) | 155 | 2 | Death or survival of patients with hepatitis; Class 1 = survived (123), 0 = died (32) |
Follow-up (window not reported) | Hospital cohort | UCI 46 |
| RISK-04 | Maternal health risk | 1,014 | 3 | Risk level assigned in the source | Current | Rural clinic monitoring, Bangladesh | Ahmed & Kashem |
| RISK-05 | Heart failure (mortality) | 299 | 2 | Death during follow-up of 4β285 days | Follow-up | Hospital cohort, Pakistan | Chicco & Jurman |
| RISK-06 | Diabetes (early detection) | 520 | 2 | Diabetes diagnosis | Current | Hospital questionnaire, Bangladesh | Islam et al. |
| RISK-07 | Hypertension (prevalent) | 4,240 | 2 | Prevalent hypertension at examination (Framingham prevalentHyp; column Risk) |
Current | Framingham Heart Study teaching dataset | HeyWhale |
| RISK-08 | Sleep disorders | 374 | 3 | Sleep-disorder category | Current | Public sleep and lifestyle dataset | Sleep Health and Lifestyle |
| RISK-09 | Diabetes (physical examination) | 1,006 | 2 | Diabetes status | Current | Hospital check-up records, China | HeyWhale |
| RISK-10 | Chronic kidney disease | 400 | 2 | Clinical CKD diagnosis | Current | Hospital caseβcontrol sample, India | Kaggle / UCI 336 |
| RISK-11 | Thyroid cancer (recurrence) | 383 | 2 | Recurrence of differentiated thyroid cancer | Follow-up (at least 10 years) | Hospital cohort, Iran | Thyroid-Dataset |
| RISK-12 | Cervical cancer | 858 | 2 | Biopsy result | Current | Hospital screening records, Venezuela | Fernandes et al. |
Regression
| ID | Task | Samples | Prediction target | Source design | Source |
|---|---|---|---|---|---|
| RISK-13 | Cardiac risk | 1,000 | 10-year ASCVD risk computed with the ACC/AHA estimator from the included inputs | Calculator output | ASCVD (Heart) Risk |
| RISK-14 | Calories | 15,000 | Energy expended during exercise | Public exercise records | Calories Burnt Prediction |
Image classification
| ID | Task | Samples | Classes | Prediction label | Source |
|---|---|---|---|---|---|
| RISK-15 | Eye diseases (six binary heads: DR, MH, DN, ODC, TSLN, normal) | 4,112 | 6 | Fundus findings (current) | Eye disease dataset |
| RISK-16 | Skin diseases | 10,015 | 7 | Dermoscopic lesion class (current) | HAM10000 / ISIC 2018 |
| RISK-17 | Chest CT | 1,000 | 4 | Normal or lung-cancer subtype (current) | Chest CT-Scan images |
Object detection
| ID | Task | Samples | Prediction label | Source |
|---|---|---|---|---|
| RISK-18 | Breast cancer | 2,358 | Lesion location and class (normal, benign, malignant); the detector returns confidence scores, not calibrated probabilities | Breast ultrasound images (BUSI) |
Task IDs match Supplementary Table 13 of the manuscript.
Notes on use in the manuscript
- Excluded predictors. RISK-05 excludes follow-up time (
time) and RISK-11 excludes post-treatment response (Response), because neither is available at prediction time. - Records used in cross-validation. Counts above refer to the source collections. Cross-validation used 835 cervical-cancer records (23 exact duplicates removed), 3,472 eye images, 988 chest CT images and 1,037 breast ultrasound images after task-specific processing.
- Fold construction. Outer folds were grouped by exact feature rows (tabular), HAM10000 lesion identifiers (skin) or normalized image paths (other images). These keys are not patient identifiers.
Dataset structure
Risk-assessment-datasets/
βββ csvfile/ # Tabular and regression data (CSV)
β βββ obesity_dataset.csv
β βββ diabetes_dataset.csv
β βββ hepatitis_dataset.csv
β βββ hypertension_risk_dataset.csv
β βββ heart_risk_dataset.csv
β βββ calories_dataset.csv
β βββ ...
βββ Img/ # Image data
βββ Breast_cancer/ # Object detection
βββ eye_disease/ # Ophthalmic
βββ Skin_Cancer/ # Dermatology
βββ chest_cancer/ # Chest CT
Changelog
- 2026-10: Added prediction-label definitions, timing and source designs for every task; corrected the hepatitis label description (mortality, not diagnosis) and the hypertension label description (prevalent hypertension); removed data files not used in the manuscript; renumbered task IDs consecutively (RISK-01 to RISK-18) to match the manuscript; removed the inaccurate "in-house" wording; corrected the citation.
Citation
@article{ye2026clinically,
title={Clinically grounded multi-agent artificial intelligence for preventive health management},
author={Ye, Dongxin and He, Sicheng and Du, Zhaowu and Yang, Yu and Xiao, Yu and Ren, Liping and Dong, Nanqing and Hu, Fang and Su, Jinsong and Zhang, Jie and Tan, Yun and Zhao, Li and Zhang, Yang and Lin, Hao},
year={2026}
}
- Downloads last month
- 4,176