DNA methylation age prediction from incomplete profiles requires evaluation under empirical and random missingness. We evaluated whole-blood chronological-age prediction using 524 development and 132 held-out GSE40279 samples and an external GSE73103 cohort (n = 355). Feature-stability analysis across 750 prespecified perturbation conditions identified a 209-CpG primary panel and a 52-CpG stringent sensitivity subset. The primary predictor combined five-nearest-neighbor reconstruction (KNN5) with Ridge regression. In the internal holdout, intact-profile prediction achieved a mean absolute error of 4.24 years, a root-mean-square error (RMSE) of 5.38 years, and Pearson r = 0.934. Under masks derived from 42 empirical templates of formalin-fixed paraffin-embedded detection failure, KNN5 reduced RMSE by 0.365 years relative to development-set mean replacement (95% bootstrap interval for the difference, -0.546 to -0.205), with larger reductions at higher missingness burdens. The reconstruction advantage persisted in complementary development analyses after randomization preserving both sample-level and CpG-level failure counts. External predictions remained associated with chronological age (r = 0.835), despite age-dependent calibration bias. Mixture analyses showed that background accuracy was central to target recovery as mixture asymmetry increased. These findings support local reconstruction within a stability-selected CpG panel for chronological-age prediction under the evaluated empirical and random missingness conditions.
