Skip to content

DATASET_FACTS.md

Latest commit

 

History

History
78 lines (61 loc) · 3.24 KB

DATASET_FACTS.md

File metadata and controls

78 lines (61 loc) · 3.24 KB

Dataset Facts

Section 1: Public Training Dataset

Field Value
Source NHANES (National Health and Nutrition Examination Survey), CDC
Years 1999–2000 through 2017–2020 (Pre-pandemic)
Total records after preprocessing ~50,000
Unique individuals ~50,000 (cross-sectional; one record per individual per survey cycle)
Age range [X–Y] years
Sex distribution [X% Female, Y% Male]
Race/ethnicity distribution Nationally representative with oversampling of Hispanic, non-Hispanic Black, non-Hispanic Asian, and older adult populations
Number of biomarkers used 42 (20 CBC + 22 CMP)
Missingness rate (mean across features) [X%]
Public availability Yes — NHANES Data

NHANES Survey Cycles Used

Cycle Included
1999–2000 Yes
2001–2002 Yes
2003–2004 Yes
2005–2006 Yes
2007–2008 Yes
2009–2010 Yes
2011–2012 Yes
2013–2014 Yes
2015–2016 Yes
2017–2018 Yes
2017–2020 (Pre-pandemic) Yes

Biomarker Sources

Training data was constructed by merging the following NHANES data files per survey cycle:

  • Laboratory — Biochemistry Profile (BIOPRO): Liver enzymes, metabolic markers, proteins
  • Laboratory — Complete Blood Count (CBC): Hematologic parameters
  • Demographics: Age (RIDAGEYR), sex, race/ethnicity

All files are publicly available as SAS Transport (XPT) format from the NHANES website.

Section 2: Internal Validation Dataset

Field Value
Source De-identified clinical blood panel records from Healome partner clinics
Total records ~1.5M blood-test records
Unique individuals [X — fill in the real number]
Average records per individual [X]
Time span [start year – end year]
Median follow-up interval [X months/years]
Geographic coverage [e.g., US — specify regions if possible]
Clinical settings [e.g., longevity/wellness clinics, primary care]
Biomarker overlap with NHANES model [X of 42] biomarkers available
Repeated measurements available Yes — [X%] of individuals have ≥2 records
Public availability No — patient privacy and governance constraints
De-identification [brief statement on method, e.g., HIPAA Safe Harbor]
Governance [brief statement — IRB, BAA, HIPAA compliance as applicable]

What I can share from the internal dataset

  • Aggregate validation metrics (see BENCHMARKS.md)
  • Summary statistics (above)
  • Short-interval test-retest variance estimates (if available)

What I cannot share

  • Individual-level data
  • Clinic-identifying information
  • Raw biomarker distributions (protected under data governance agreements)

Why this matters: This document exists so that anyone evaluating this work can understand exactly what data was used, how it was collected, and what constraints apply to the validation claims. A researcher reading this should immediately understand what data is available, what has been done with it, and what constraints apply.