Multisite collaborations and large databases in psychiatric neuroimaging: advantages, problems, and challenges.

Studying large numbers of subjects in psychiatric neuroimaging research is often necessary to generate sufficient statistical power to reveal more subtle differences missed insmalleranalysesandtoallowsubtypinginwhatareoften heterogeneous diagnostic groups. Such an approach is probably essential in the case of clinical studies dealing with groups that are difficult to recruit from a single site in large numbers. This issue is brought into even sharper relief where neuroimaging and genetics are both part of a particular study. Dealing with such high-dimensional datasets often implies the need for both multisite collaborations with regard to data collection and specialized statistical approaches with regard to data analysis. These issues deserve detailed discussion and careful examination as multisite imaging collaborations begin to proliferate in neuropsychiatric research. Recent large-scale multisite collaborations with neuroimaging as an essential component include the structural and functional imaging arms of the Biomedical Informatics Research Network (BIRN, http://www.nbirn.net/), the Alzheimer’s Disease Neuroimaging Initiative (ADNI, http://www. adni-info.org/), the Bipolar-Schizophrenia Network on Intermediate Phenotypes (B-SNIP, http://www.b-snip. org/), and The North American Prodrome Longitudinal Study (NAPLS, http://www.schizophreniaforum.org/ new/detail.asp?id = 1339), all funded by National Institutes of Health. As noted above, there are many pluses of multisite collaborations. Results gathered in diverse clinical populations are more likely to generate generalizable data and to provide adequately sized and powered samples. Planning such collaborations forces the issue of standardization of imaging collection and processing and the creation of centralized neuroinformatics repositories. This in itself is complex and probably one of the major reasons brain imaging is lagging behind large-scale genetic studies. It is likely that multisite collaborations benefit by being vetted by a team of collaborating scientists from the beginning, although in the minus column they require more administrative oversight/infrastructure to ensure all aspects of the study are performed consistently. Naturally, this also leads to tool development and dissemination (which is one of the goals of Functional BIRN). Collections of large datasets inevitably lead to attempts to automate and increase efficiency of associated procedures including automated software to read image data and radiological header files and automatically report back departures from protocol. Each site must remain vigilant to fluctuations in image quality and stability not only day to day but also over longer time periods in terms of signal to noise. For example, one multisite study discovered functional imaging fluctuations, ultimately traced after a heroic search to the occasional passing of metro trains in a tunnel adjacent to the hospital. Other problems associated with multisite collaborations include the fact that use of different platforms of neuroimaging equipment and other related site differences add undesirable heterogeneity, necessitating complex preparation in terms of sending both inert radiologic phantoms and traveling subjects, often the investigators, in the form of ‘‘human phantoms’’ from site to site to ensure maximum comparability, a strategy used both by BIRN and ADNI. The ADNI study, that involves over 100 participating sites in the United States, devoted much time and effort to the preparatory comparison of different magnetic resonance imaging (MRI) scanning platforms in order to obtain maximally comparable structural imaging sequences on different model scanners. This also involved working closely with the scanner manufacturers and obtaining their collaboration in the face of often closely guarded proprietary information. Despite such precautions, multisite studies still need to add site as a covariate in analyses, advanced analysis approaches such as independent component analysis (ICA) employed in the Kim article in this issue (and see Meda et al 2008). The former approach may be able to filter out much of the multisite heterogeneity. However, while such approaches can be used to identify and to minimize site effects, they cannot fully remove them. Experience from the BIRN study identifies another key element in the group comparison approach. Calhoun Vince 2008 (personal communication) found in most cases that between-group differences were consistent across site (few interactions of site and group) but that there were very significant main effect site differences. Schizophrenia Bulletin vol. 35 no. 1 pp. 1–2, 2009 doi:10.1093/schbul/sbn166 Advance Access publication on November 20, 2008