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

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 generalise across scanning sites. In this study, we first establish best performing within-site models, and then systematically investigate cross-site generalisation in first-episode schizophrenia (FES) classification and evaluate strategies for mitigating site-related distribution shifts. Using data from two acquisition sites (n = 389 in total), we perform train-on-site/test-on-site experiments to analyze performance degradation under domain shift and examine the effectiveness of ComBat, optimal transport, and adversarial adaptation strategies. Across functional, structural, and diffusion-based features, both traditional machine learning (TML) and neural network (NN) models achieve 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 degrades substantially across all approaches, highlighting the impact of site-related variability. Distribution-alignment methods partially mitigate this degradation, with ComBat and optimal transport yielding more consistent cross-site improvements than adversarial adaptation. Increasing model complexity alone does not result in systematic performance gains, and simple models combined with effective alignment strategies often perform comparably to more complex neural architectures, while multimodal feature fusion does not consistently outperform functional connectivity alone. Overall, our findings indicate that controlling for site effects is more critical than model complexity for achieving generalisable classification in FES, underscoring the importance of rigorous evaluation designs and explicit distribution-alignment strategies for neuroimaging-based predictive models with potential clinical utility.