Supplementary Figure 4: Comparison of models trained within site to models trained across site.
From: Common brain disorders are associated with heritable patterns of apparent aging of the brain
In the test sample (n=10,141 independent subjects), for each diagnosis and scanner, we trained a machine learning model on data from all available healthy controls acquired at a given scanner and predicted brain age on data from all cases collected at the respective scanner. Predicted brain age in cases from the within-site models correlated significantly with predicted brain age from the main models (mid column). Likewise, the resulting brain age gaps (right column) were significantly correlated between within-site models (accounted for age, age², sex and Euler number) and main models (accounted for age, age², sex, Euler number and scanning site), indicating that scanning-site independent models provide similar estimates of apparent aging patterns as the models built on cross-site imaging data. The figure reports Pearson r with two-sided p values, all P<FDR (Benjamini-Hochberg). Model performance (left column) is higher in the main models that use more data (see also Supplementary Fig. 5), and we have therefore used the across-site predictions for the main analysis.
