ADM Prediction (Made Before Running Models)
Predicted winner: Hybrid. KDM provides an age-structured statistical prior using biomarker-age relationships. A learned model must estimate those patterns from only 1,907 samples. KDM is largely linear and does not represent all interactions; the hybrid uses its estimate as a feature and learns residual nonlinearities.
Expected: KDM > ML alone (insufficient data); Hybrid > KDM (residual nonlinearities). Actual: KDM r=0.527, ML r=0.478, Hybrid r=0.631. Prediction confirmed.
Three grouped bar clusters showing Pearson r for KDM (domain), ML, and Hybrid across forward split-half, reverse split-half, and self-rated-health correlation. The hybrid has the highest forward split-half correlation. The learned model has the highest reverse and self-rated-health correlations. These are internal checks, not external clinical validation.
The KDM formula regresses each biomarker against chronological age to get slopes and residual standard deviations, then combines them using a weighted maximum-likelihood estimator. The result is a biological age estimate depending only on blood biomarkers.
Split-half validation splits the sample in two halves. KDM coefficients are trained on Half A, predictions made for Half B. Then we check: does the predicted biological age correlate with aging outcomes in Half B? This guards against the circular reasoning that inflated earlier studies (r=0.965 was just predicting chronological age).
Self-rated health is an independent criterion within the same NHANES cohort: people who rate their health as “Poor” should have higher biological age deltas than those who rate it “Excellent.” It is not an external cohort or an objective clinical endpoint.
ADM Insight
This investigation had to be rewritten. The original hybrid model achieved r=0.965 — suspiciously perfect. It was predicting risk factors from biomarkers used to define those risk factors. Split-half validation reduces that leakage: train the estimator in one half and evaluate its associations in the other. The hybrid correlation is 0.631, not 0.965. KDM_Delta is the highest-importance feature in the hybrid model, with 44% of impurity-based feature importance; that is predictive importance, not a causal contribution. Domain knowledge isn't replaced by ML. It's the backbone that ML extends.
Why does ML alone (r=0.478) underperform KDM (r=0.527)? With only 1,907 samples and 7 biomarkers, the learned-only model has less data from which to estimate stable patterns. KDM supplies an age-structured biomarker summary, and the hybrid adds learned residual structure. This is the ADM thesis in miniature: encode a defensible prior, then test whether data-driven residuals improve held-out results.
Small sample: N=1,907 from NHANES 2017-2018. This limits estimation of complex patterns. A larger cohort is needed to test whether the model ranking persists.
Wide individual dispersion: KDM biological-age deltas have a standard deviation of 28.06 years in this sample. That spread makes the estimates unsuitable as precise individual ages; the analysis supports comparison of held-out associations, not a personal diagnostic reading.
Cross-sectional only: Biological age is estimated from a single snapshot. Longitudinal tracking (does higher bio-age predict faster decline?) would be a stronger validation than cross-sectional correlation.
Self-rated health is not external validation: SRH is an independent criterion in the same cohort, but it is subjective and associated with depression, socioeconomic status, and cultural factors. Stronger validation would use an external cohort and longitudinal endpoints such as mortality, disease onset, or functional decline.
Seven biomarkers only: KDM uses 7 blood biomarkers. The result does not establish how the models would perform with epigenetic, proteomic, or other aging measures.
No fairness analysis: Bio-age estimation may perform differently across racial/ethnic groups, sexes, and socioeconomic strata. Subgroup validation not yet performed.