Rohithreddybc/rised-healthcare-eval-dataset

RISED Synthetic Clinical Cohort (10,000 patients)

A fully synthetic adult clinical cohort generated deterministically (random_state = 42) by a Synthea-inspired computational model implemented in the `rised` Python package. No real patient records were used at any stage. The cohort is intended as a methodological testbed for the RISED Framework and is demographically heterogeneous to support subgroup-level evaluation.

This is the reference dataset used in the demonstration application of the RISED Framework — a five-dimension pre-deployment evaluation framework for clinical AI decision-support systems (Reliability, Inclusivity, Sensitivity, Equity, Deployability).

Dataset Summary

Demographic Composition

Mean Charlson Comorbidity Index: 0.99 ± 1.20 (full); 1.86 ± 1.33 (outcome=1).

Features

Demographics: age, sex_male, age_group, sex, race, insurance, ins_medicare, ins_medicaid, ins_private (insurance-type indicators included as a demographic axis on which to evaluate subgroup performance)

Clinical: cci_score, has_hypertension, has_diabetes, has_chf, has_ckd, has_copd, has_mi, has_cvd, has_dementia, has_cancer, has_metastatic, prior_hosp_count, ed_visits_count, bmi

Neighborhood: adi_score (deprivation index, 1–100 scale)

Outcome: high_need (binary; 1 = top-30% derived clinical risk score). The column is named high_need for backward compatibility with earlier versions of the codebase; it represents a generic adverse-clinical-outcome label and the cohort is not specific to any particular clinical use case or deployed risk-stratification program.

Outcome Definition

The binary outcome label is derived from a logistic transformation of age, diabetes, congestive heart failure, chronic kidney disease, COPD, prior myocardial infarction, CCI, prior hospitalization count, ED utilization count, and the deprivation index, with additive Gaussian noise (σ=0.5). Patients in the top 30% of the predicted score receive label = 1.

Important: Because the outcome is derived directly from the feature space, this dataset is suitable for evaluation framework demonstrations but not for benchmarking model accuracy on a real-world prediction task. Real EHR cohorts introduce distribution shifts and access-barrier distortions absent from synthetic data.

Intended Use

• —Primary: Demonstrating the RISED Framework for pre-deployment evaluation of clinical AI decision-support systems.
• —Secondary: Teaching, methodological development, and reproducibility benchmarking for fairness, calibration, and sensitivity tooling.

Not intended for: training production clinical models, benchmarking discrimination performance against real-world systems, or any deployed clinical use.

Usage

from datasets import load_dataset

ds = load_dataset("Rohithreddybc/rised-synthetic-cohort-10k")
df = ds["train"].to_pandas()
print(df.shape)         # (10000, 26)
print(df["high_need"].mean())  # 0.30

Or load directly with pandas:

import pandas as pd
df = pd.read_csv(
    "hf://datasets/Rohithreddybc/rised-synthetic-cohort-10k/synthetic_cohort_10k.csv"
)

To regenerate from source (deterministic):

from rised.datasets import generate_synthea_cohort
df = generate_synthea_cohort(n=10000, random_state=42)

Citation

If you use this dataset, please cite the accompanying paper:

@article{bellibatlu2026rised,
  author  = {Bellibatlu, Rohith Reddy},
  title   = {{RISED}: A Pre-Deployment Evaluation Framework for Clinical {AI}
             Decision-Support Systems Spanning Reliability, Inclusivity,
             Sensitivity, Equity, and Deployability},
  year    = {2026},
  journal = {Artificial Intelligence in Medicine (under review)},
  url     = {https://github.com/rohithreddybc/rised-healthcare-eval}
}

License

MIT. The dataset is fully synthetic and contains no information derived from real patients; redistribution and derivative works are unrestricted under MIT terms.

Limitations

1. 1.Synthetic only. Distributions reflect a generative model, not real-world epidemiology. Results obtained on this cohort do not generalize to real clinical populations without further validation.
2. 2.Self-derived outcome. The outcome label is a function of the feature space, so high accuracy is expected and does not indicate real predictive skill. Use for methodology evaluation only.
3. 3.Single random seed. All values are deterministic at seed = 42; future versions may include alternative seeds.

Contact

Rohith Reddy Bellibatlu — rohithreddybc@gmail.com — ORCID: 0009-0003-6083-0364