Contributions of Site Heterogeneity in Multi-site Brain-Phenotype Predictions

Age prediction from harmonized structural MRI shows pooled performance (r = 0.73) higher than full multi-site training (r = 0.63) across 3486 participants from 9 sites.

• Why it matters: Single Pearson correlations can mask how training-set size, feature/outcome variability, and true between-site effect heterogeneity shape machine learning brain-phenotype prediction.
• What they did: Using Reproducible Brain Charts structural MRI, the authors compared within-site (Local), size-downsampled multi-site (Multi-Down), full multi-site (Multi-Full), and pooled (Overall) training while including the target site in both training and testing.
• The result: Between-site heterogeneity was extensive (age I^2 ≈ 95%), and pooled gains reflected mechanical inflation from larger outcome variance; at matched sample size, within-site training outperformed harmonized multi-site, with a similar pattern for brain-p-factor signals.

AI summary of the abstract · may contain errors · not medical advice · read the paper.