Extended Data Fig. 1: Accuracy and reliability of brain age estimation models across tissue classes and samples.
(a) grey matter, (b) white matter, and (c) combined grey and white matter age estimation models. Blue dots in the first three plots (from left to right) show brain-predicted age estimates plotted against the chronological age in the UKB discovery sample (n = 32,634; white-British ancestry), UKB replication sample (n = 21,881; multi-ancestry), and LIFE-Adult replication sample (n = 1,833; European ancestry). To facilitate comparisons, results of the UKB discovery sample are also shown as grey dots in the background of the LIFE replication plots. At this stage, brain-predicted age estimates have not yet been bias-corrected for regression dilution, that is, younger participants’ ages are systematically overestimated and vice versa, as indicated by the linear regression line (solid) crossing the identity line (dashed). The fourth plot shows the test-retest reliabilities of brain age gap (i.e., the difference between brain-predicted and chronological age) in a subset of the UKB discovery (grey dots, n = 3,751) and UKB replication sample (blue dots, n = 395). For test-retest comparisons, brain age gap was residualized for effects of age, age2, sex, scanner site, and total intracranial volume. MAE: mean absolute error; rho: product-moment correlation coefficient. ICC: intraclass correlation coefficient (C, 1).
