Pooling cohorts is a stated premise of several CNS projects — training on one dataset and validating on another, or combining cohorts to reach a usable sample size. The obstacle is that scanner and site leave a signature on the data that is often larger than the clinical effect being studied. Different vendors, field strengths, head coils, and acquisition parameters all shift derived measures like cortical thickness or connectivity strength.
The danger is not merely added noise. If diagnosis is unevenly distributed across sites — and it usually is, because each site recruits its own participants — then site is a confound, and a classifier can reach excellent accuracy by learning which scanner produced the image rather than anything about the brain. That failure mode is common enough that apparent cross-cohort success should always prompt the question of whether site was separable.
ComBat, borrowed from genomics batch correction, is the most widely used answer: model site as a batch effect and remove it while preserving the biological covariates you name. Variants extend it to nonlinear age effects and to longitudinal data. It is not magic — harmonization cannot recover information an acquisition never captured, and removing site variance will also remove real biological variance that happens to align with site.
Related¶
Mixed models — site as a random effect
Data leakage — site leakage across a train/test split
Neurobagel and semantic annotation — harmonizing what variables mean
Transfer learning — the other approach to a domain shift