Compact DNA methylation panels for forensic age estimation: cross-cohort evaluation and calibration using public whole-blood data

DNA methylation-based age estimation is promising for forensic intelligence, but transferability across cohorts and platforms remains a major challenge. We evaluated whether compact, literature-supported blood CpG panels retain predictive performance across independent public cohorts and whether limited marker expansion or local calibration improves accuracy. A health-filtered GSE207605-derived resource was used for model development after excluding GSE40279 and GSE69270, leaving 8,986 samples for nested group-aware cross-validation. Three Elastic Net models containing four, eight, or nine CpGs were compared and externally evaluated in GSE40279 (n = 656), GSE69270 (n = 184), and GSE141682 (n = 42). Internal mean absolute error was 5.97, 5.86, and 5.84 years for the four-, eight-, and nine-CpG models, respectively. Raw external performance was cohort dependent: the four-CpG model achieved the lowest MAE in GSE40279 (8.62 years) and GSE69270 (3.90 years), whereas the eight-CpG model performed marginally better in GSE141682 (7.85 years). In leakage-free repeated-split calibration, intercept-only adjustment reduced the eight-CpG model mean held-out MAE from 8.83 to 5.56 years, 4.28 to 2.46 years, and 7.77 to 3.92 years, respectively. These findings support compact blood-based panels as the primary forensic strategy and indicate that cohort-specific calibration may be more consequential than adding markers. Prospective laboratory validation remains required before casework implementation.