Age-range dependence and prediction compression in blood-based chronological-age models

The dependence of test-set R² on the variation of chronological age is established, but blood-age models are often compared without their evaluation age distributions, absolute errors, or simple baselines. We examined these reporting issues in routine blood-based age prediction and compared the age spans and R² values reported by earlier studies. We analyzed 17 routine blood biomarkers in a Saudi laboratory cohort of 18,651 adults aged 20–84 years and an independent NHANES 2017–2018 cohort of 2,067 adults aged 20–79 years. Random Forest models were refitted and assessed with five-fold out-of-fold prediction within broad and restricted age windows, alongside training-fold mean-age and sex-only baselines. We assessed prediction-range slopes, biomarker variance components, and model-class robustness. A targeted, non-systematic comparison extracted reported age spans and held-out R² from four primary blood-based age-model reports. Saudi and NHANES broad-range R² values were 0.382 and 0.378, respectively. At ages 30–74 years they were 0.278 and 0.280, and at ages 40–64 years they were 0.134 and 0.074. Within individual decades, R² was near zero or negative. At ages 40–64, the model reduced mean absolute error relative to the sex-only baseline by 0.51 years in Saudi and 0.30 years in NHANES. Predicted-age slopes were 0.356 and 0.337. Published reports showed differing R² values even at similar age spans and a non-monotonic pattern across separately fitted age strata; the literature comparison therefore cannot assign between-study differences to age width. Broad-range R² can coexist with compressed predictions and little added age discrimination within narrow windows. The age-variance dependence of R² is not a new statistical finding; these data show its practical reporting consequences for a defined biomarker panel. Studies should report age distribution, absolute error, baselines, validation design, and prediction-range slope alongside R². Chronological-age fit alone does not validate biological aging.