Domain Adaptation Improves Cross-site Classification of First-episode Schizophrenia from Multimodal Neuroimaging Data

Resting-state fMRI functional connectivity yields the most robust unimodal features for first-episode schizophrenia classification, while cross-site transfer across two sites degrades balanced accuracy by 5–12 points.

• Why it matters: Neuroimaging classifiers often fail to generalize across scanning sites, limiting clinical translation; this study targets site-related distribution shifts in first-episode schizophrenia (FES).
• What they did: Using data from two acquisition sites (n=544), the authors trained within-site TML and NN models and then tested ComBat, CovBat, optimal transport, and deep adaptation for cross-site FES classification.
• The result: Distribution-alignment partially mitigated cross-site degradation; harmonization recovered 6–8 points for functional connectivity in one transfer direction, while deep adaptation recovered similar gains less consistently, and multimodal fusion did not consistently beat functional connectivity alone.