There has been a growing interest in the use of resting-state functional connectivity (FC) to quantitatively examine the neural mechanisms of mental disorders. In particular, the application of various machine-learning techniques has enabled the data-driven identification of disorder-specific patterns of FC, the use of which has led to the automatic case-control classification of individuals with an accuracy of 80-90%. However, none of the previous classification methods have ever been successfully validated for an independent cohort due to overfitting and interference from nuisance variables (NVs), thus hampering their clinical application. Here, using a multiple-site dataset from Japan, we developed a classifier for autism spectrum disorder (ASD) by focusing on abnormal FCs in ASD as revealed by resting-state functional connectivity magnetic resonance imaging (rs-fcMRI). To overcome the difficulties associated with overfitting and NVs, we developed a novel machine-learning algorithm that automatically and objectively identified a small number of abnormal FCs in ASD (0.2% of all FCs considered). This classifier showed high-level diagnostic accuracy for the individuals in the dataset, and also demonstrated a marked degree of generalization for an independent cohort comprised of a different ethnic population in the USA. The same set of FCs in the classifier accurately predicted the communication domain score of the standard diagnostic instrument. Thus, we have established a reliable rs-fcMRI-based biomarker for ASD that reveals a direct link between the underlying neural mechanisms and behavioral characteristics associated with ASD. Finally, we examined the specificity of the ASD classifier for ASD by investigating its generalizability to other psychiatric disorders. We found that our classifier did not distinguish individuals with major depressive disorder or attention-deficit hyperactivity disorder from their controls, but that it did have a moderate ability to distinguish patients with schizophrenia from their controls. Therefore, our findings support the possibility that exploring neuroimaging-based dimensions to quantify multiple-disorder spectrum may contribute to more biologically oriented diagnostic systems in clinical psychiatry.
 <Author's abstract>