NHANES Epigenetic Clock Mortality Validation

A small, reproducible pipeline that asks one question with public data:

Do DNA-methylation "aging clocks" predict death better than chronological age and sex alone?

It links eight epigenetic clocks measured on NHANES 1999–2002 participants (aged 50+) to ~17–20 years of mortality follow-up from the NCHS public-use Linked Mortality Files, fits Cox proportional-hazards models, and measures how much each clock improves out-of-sample discrimination (test C-index) over a plain age + sex baseline.

Cohort: n = 2,532 eligible participants aged 50+ with both DNAm biomarkers and mortality linkage. 1,361 deaths (53.8%), median follow-up 17.1 years.

Headline result

	C-index (test)	Δ vs. age+sex	HR per SD (95% CI)	p
Age + sex (baseline)	0.746	—	—	—
+ GrimAge2Mort	0.769	+0.0235	2.06 (1.84–2.30)	2.5e-36
+ GrimAgeMort	0.766	+0.0202	2.09 (1.86–2.37)	7.0e-33
+ DunedinPoAm	0.757	+0.0116	1.36 (1.27–1.45)	1.2e-20
+ 5 age-trained clocks	0.747–0.750	+0.0016 … +0.0044	—	—

Mortality-trained clocks (GrimAge2, GrimAge) carry real, highly significant signal beyond age and sex, but the absolute discrimination gain is modest (≤0.024 C-index). Full numbers in results/cindex_comparison.csv and RESULTS.md.

Pipeline (four steps)

download  →  verify  →  build_cohort  →  analysis

1. download (src/download.py) — documents the exact data sources and validates that the required raw files are present in data/raw/.
2. verify (src/verify_variables.py) — loads the DNAm files and confirms the clock column names/labels and that SEQN/WTDN4YR exist.
3. build_cohort (src/build_cohort.py) — merges DNAm + demographics + mortality on SEQN, applies explicit eligibility filters, constructs survival variables (time_years, event), z-scores predictors, and writes a stratified 70/30 train/test split (random_state=42).
4. analysis (src/analysis.py) — fits Cox models on the train split, evaluates Harrell's C-index on the held-out test split, checks the proportional-hazards assumption, and writes the results table and figures.

Run it (one command)

python -m venv venv && source venv/bin/activate
pip install -r requirements.txt

# place raw NHANES files in data/raw/ first (see Data sources below), then:
python run_all.py        # or:  make all

Outputs:

• results/cindex_comparison.csv — per-clock C-index, Δ, HR, 95% CI, p
• results/ph_assumption_check.txt — PH diagnostics for base + best model
• figures/incremental_cindex.png, figures/km_by_grimage_tertile.png

Raw NHANES files are not committed (NCHS does not permit redistribution, and the mortality files require agreeing to a data-use agreement). The derived analytic_cohort.csv is also not committed — it is regenerated from the raw files. The small, deterministic data/processed/train_test_split.json is kept for exact reproducibility, but run_all.py regenerates an identical split from seed 42 if it is absent.

Data sources (exact URLs used)

NHANES DNA Methylation Epigenetic Biomarkers (combined 1999–2002 cycles) → data/raw/dnmepi.sas7bdat https://wwwn.cdc.gov/nchs/data/nhanes/dnam/dnmepi.sas7bdat (landing page: https://wwwn.cdc.gov/nchs/nhanes/dnam/)

NHANES Demographics

• 1999–2000 → data/raw/DEMO_1999.xpt — https://wwwn.cdc.gov/Nchs/Data/Nhanes/Public/1999/DataFiles/DEMO.xpt
• 2001–2002 → data/raw/DEMO_2001.xpt — https://wwwn.cdc.gov/Nchs/Data/Nhanes/Public/2001/DataFiles/DEMO_B.xpt

NCHS Linked Mortality Files (public-use, 2019 linkage)

• 1999–2000 → data/raw/NHANES_1999_2000_MORT_2019_PUBLIC.dat
• 2001–2002 → data/raw/NHANES_2001_2002_MORT_2019_PUBLIC.dat
• https://ftp.cdc.gov/pub/Health_Statistics/NCHS/datalinkage/linked_mortality/ (info: https://www.cdc.gov/nchs/data-linkage/mortality-public.htm)

Clocks used (verified column names): HorvathAge, HannumAge, SkinBloodAge, PhenoAge, GrimAgeMort, GrimAge2Mort, DunedinPoAm, HorvathTelo. Key variables: SEQN, RIDAGEYR, RIAGENDR, MORTSTAT, PERMTH_EXM, ELIGSTAT.

Limitations

• Unweighted. These models ignore the NHANES complex survey design, so the estimates are not nationally representative — they describe this analytic sample only. The design variables (WTDN4YR, SDMVPSU, SDMVSTRA) are kept in the cohort so anyone wanting survey-weighted estimates can produce them.
• Public-use mortality file. Follow-up time and cause of death are intentionally perturbed in the public-use Linked Mortality Files to protect confidentiality; the restricted-use files are more precise.
• Single cohort, no external validation. Train/test are two splits of the same NHANES sample. There is no replication in an independent cohort, so the C-index gains should be read as in-sample-cohort, not externally validated.
• Proportional hazards. The age + sex baseline satisfies PH, but in the best clock model both age and the clock show mild PH violations (large n makes the test sensitive). These are reported, not corrected away.
• Discrimination only. This validates ranking (C-index) and hazard ratios, not calibration or absolute risk.

Repository layout

.
├── run_all.py            # one-command pipeline
├── Makefile              # make all / install / build / analysis / clean
├── src/
│   ├── download.py       # document sources + validate raw files
│   ├── verify_variables.py
│   ├── build_cohort.py   # merge + filter + split
│   └── analysis.py       # Cox models, C-index, PH checks, figures
├── data/
│   ├── raw/              # NHANES files (not committed; download yourself)
│   └── processed/        # train_test_split.json (cohort CSV is gitignored)
├── results/              # cindex_comparison.csv, ph_assumption_check.txt
├── figures/              # incremental_cindex.png, km_by_grimage_tertile.png
├── RESULTS.md
├── CITATION.md
└── LICENSE

License & credit

Code is MIT-licensed (LICENSE). The data is produced by the U.S. National Center for Health Statistics and is subject to NCHS terms of use; please credit NHANES and the NCHS Linked Mortality Files — see CITATION.md.