Existing approaches to estimate and characterize whole brain time-varying connectivity from fMRI data have shown considerable promise, with exponential growth in research in this field. We and others have developed a powerful set of tools that are now in wide use in the community. However, the impact of mental illness on brain connectivity is complex, and as we show, limitations in existing methods often result in missing important features associated with brain disorders (e.g. transient fractionation of the spatial structure of brain networks). Some of these important limitations include 1) the most widely-used approaches often require a number of prior and limiting assumptions that are not well studied, 2) methods often assume linear relationships either within or between networks over time, and 3) methods assume spatially fixed nodes and ignore the possibility of spatially fluid evolution of networks over time. We propose a novel family of models that builds on the well-structured framework of joint blind source separation to capture a more complete characterization of (potentially nonlinear) spatio-temporal dynamics while providing a way to relax other limiting assumptions. Our models will also produce a rich set of metrics to characterize the available dynamics and enable in depth comparison with currently avail- able models including those that are model based. We will extensively validate our approaches in a variety of ways including simulations and evaluation of rigor and robustness in large normative data sets. Finally, we will apply the developed tools to study the important area of dynamic properties in mental illnesses including schiz- ophrenia, bipolar disorder, and the autism spectrum. There is considerable evidence of disruption of dynamics in all three disorders, and as we show the use of static (or even exiting dynamic) approaches can miss important information about brain related differences associated with each. We will provide open source tools and release data throughout the duration of the project via a web portal and the NITRC repository, hence enabling other investigators to use our approaches and compare their own methods with our own. Our tools have wide appli- cation to the study of the healthy brain as well as many other diseases such as Alzheimer's disease and attention deficit hyperactivity disorder. 38

Public Health Relevance

There is considerable interest in approaches that capture time-varying connectivity, however existing ap- proaches suffer from three major limitations: 1) extraction of networks and estimation of dynamic features for the most widely used approaches typically makes many limiting assumptions and rely on user defined parameters whose optimality is not guaranteed (e.g. choice of window size, filter settings), 2) methods typically assume relatively simple linear relationships, and 3) most approaches assume the spatial nodes are fixed in time, effec- tively ignoring the possibility of functionally fluid nodes. We propose to address these limitations by developing a family of flexible spatio-temporal models that leverage and blend the concepts of joint blind source separation and deep learning, which we show can reveal important information missed by existing models. We will apply the proposed models to study spatial and temporal aspects of dynamic connectivity across mental illness includ- ing schizophrenia, bipolar disorder, and autism spectrum. 37

Agency
National Institute of Health (NIH)
Institute
National Institute of Mental Health (NIMH)
Type
Research Project (R01)
Project #
1R01MH123610-01A1
Application #
10156006
Study Section
Special Emphasis Panel (ZRG1)
Program Officer
Ferrante, Michele
Project Start
2021-03-19
Project End
2026-01-31
Budget Start
2021-03-19
Budget End
2022-01-31
Support Year
1
Fiscal Year
2021
Total Cost
Indirect Cost
Name
Georgia State University
Department
Miscellaneous
Type
Organized Research Units
DUNS #
837322494
City
Atlanta
State
GA
Country
United States
Zip Code
30302