In the same study, those who would later develop schizophrenia (“converters”) could be distinguished from the nonconverters on the basis

of smaller gray matter volume mainly in limbic and temporal areas. These findings may support biological models positing progressive cortical volume loss as a risk factor ATM inhibitor for schizophrenia (Wood et al., 2008). Biomarkers derived from pattern classification do not come with clear cut-off points and depend strongly on the experimental parameters (e.g., numbers of scanned voxels) and analytical approches (e.g., the algorithm used for feature selection), and their practical relevance therefore needs to be demonstrated in multicenter studies, where the prediction accuracy of a template derived from one scanner is tested in data sets from others (Klöppel et al., 2008). Such confirmation in independent test samples is also needed to overcome doubts about the prediction estimates obtained through cross-validation in small samples (Isaksson et al., 2008). However, based on the promising results obtained so far, it can reasonably be expected that pattern classification of brain imaging data, in combination with clinical and psychometric data, will improve our ability to predict the course of psychiatric diseases. Although the reliability of structural

imaging measures is high on the same scanner, it is insufficient when tested IOX1 cost across scanners (Kruggel et al., 2010). However, improvements are to be expected

from wider use of high-field scanners with better image quality and segmentation results. Replication is also likely to be better if at least the field strength is kept constant across sites. The successful discrimination of AD patients from controls in a multicenter study of structural imaging data is promising in this respect (Klöppel et al., 2008). Less information is available about the reliability of specific functional imaging measures, because these would many in principle have to be computed for each individual cognitive paradigm. The literature on reproducibility of task-related activation converges to report consistency in the qualitative activation patterns, but considerable intraindividual variability, across scanning sessions (Gountouna et al., 2010) (Table 1). We are thus still far away from fMRI-based biomarkers at the individual level. The situation is similar for resting state measures, which have been too heterogeneous across individuals to allow for the development of stable biomarkers of disease (Greicius, 2008). However, recent work on graph theoretical metrics of functional connectivity has yielded promising results for the intraindividual reliability of some metrics (Braun et al., 2012). A first step toward the development of biomarkers from resting state activation metrics would thus be to achieve agreement on standardized analysis procedures based on the measures with the highest reproducibility.