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

Abstract

Identifying robust neuroimaging markers associated with schizophrenia is essential for advancing research and informing clinical understanding. However, a major obstacle to clinical translation is the limited ability of neuroimaging-based classification models to generalize across scanning sites. In this study, we first establish best performing within-site models, and then systematically investigate cross-site generalization in first-episode schizophrenia (FES) classification and evaluate strategies for mitigating site-related distribution shifts. Using data from two acquisition sites (n=544), we examine ComBat, CovBat, optimal transport, and deep adaptation. Across functional, structural, and diffusion-based features, both traditional machine learning (TML) and neural network (NN) models achieved comparable performance in within-site classification, with resting state fMRI functional connectivity providing the most robust unimodal features. When models are transferred across sites, performance degraded substantially across all approaches, by 5 to 12 percentage points of balanced accuracy. Distribution-alignment methods partially mitigated this degradation. Harmonization recovered 6 to 8 points for functional connectivity in one transfer direction, mainly by correcting a displaced decision threshold. Deep adaptation recovered similar gains but less consistently. Increasing model complexity alone did not result in systematic performance gains. Across the classifiers tested on this pair of sites, multimodal feature fusion did not consistently outperform functional connectivity alone. Overall, our findings indicate that controlling for site effects was more critical than model complexity for achieving generalizable classification in FES. These results come from two sites of a single coordinated study, and whether they hold for more heterogeneous sites remains to be tested.