Neuropsychiatric disorders are particularly serious health issues that affect the cognitive ability, emotion and behavior of individuals. The increasing prevalence and severe symptoms reflect the urgent need for effective diagnostic classification methods for neuropsychiatric disorders. Recent studies combine multi-site functional magnetic resonance imaging (fMRI) data and domain adaptation techniques to address this problem. However, most related studies primarily focus on feature matching across the source and target domains, neglecting the issue of negative transfer brought by source data that exhibit low relevance with the target domain. In this study, we propose a two-stage instance selection domain adaptation algorithm (TS-ISDA) to tackle this issue. Specifically, we propose the minimum transfer distance (MTD) for direct distance measurement between a given sample and distribution. The MTD-based source instance selection and the target instance refinement steps are performed progressively to filter out unrelated source data and target instances with low-quality pseudo labels to ensure the effectiveness of the category-level feature alignment process. Then the local maximum mean discrepancy between the selected source and target data is calculated as the domain alignment loss. We conduct cross-site diagnostic classification studies of schizophrenia (SCZ), major depressive disorder (MDD) and autism spectrum disorder (ASD) based on fMRI data of 1003, 1043 and 607 participants from six sites, respectively. TS-ISDA achieves the highest average accuracies of 81.2
