CRCNS: Bayesian Analysis of Neural-Behavioral Interactions in Mental Illness

点击翻译按钮获取中文摘要

DESCRIPTION (provided by applicant): Schizophrenia is an illness with enormous public health significance, affecting approximately 1% of the population and inflicting immense personal and economic cost. Treatment possibilities for schizophrenia are still limited, at least in part because of poor understanding of its anatomical, neural, cognitive, and genetic substrates. Morphometric and functional neuroimaging technologies suggest that schizophrenia affects distributed brain circuits, but identifying significant circuits in such rich data sources is challenging. The primary goal of this proposed project is to identify key functional and anatomical networks that are altered in schizophrenia. To accomplish this, we will develop novel, data-driven Bayesian computational model search techniques that can automatically locate significant and clinically relevant network descriptions of rich, multimodal schizophrenia data. These network models will inform us about the specific neural and mental substrates of schizophrenia, will correlate them with exogenous clinical assessments and treatment outcomes, and will ultimately guide both future investigations of schizophrenia and treatment courses. Data will be obtained from the Clinical Imaging Consortium of the Mental Illness and Neuroscience Discovery (MIND) Institute, who are performing an unprecedented multi-site, multi-modality study of schizophrenia. This study will examine hundreds of schizophrenic patients and a matched number of controls by collecting a sophisticated suite of neuroimaging data (including structural MRI, fMRI, DTI, EEG and MEG data) and genetic, clinical and psychiatric variables from each subject. We will use dynamic Bayesian networks (DBNs) as our model class and DBN structure search methods as the statistical model induction method. We will couple these methods to rigorous confidence testing, controls for multiple hypothesis evaluation, and expert evaluation of the resulting models. The advantages of this approach are that it can identify a wider class of relationships than can linear techniques; the resulting models have a straightforward interpretation as activity networks; it allows the incorporation of domain knowledge as Bayesian structural priors; and it can be naturally extended to incorporate exogenous variables or alternate imaging modalities. Our work will yield novel understanding of the neural network substrates of schizophrenia and their relationships to behavioral, genetic, and clinical features and treatment outcomes. This work will contribute toward improved diagnostics and therapies for schizophrenia, a disease that affects millions of people.