Multivariate methods for identifying multitask/multimodal brain imaging biomarkers

Georgia State University, Atlanta GA

Investigators

Linked publications & trials

Abstract

Project Summary/Abstract The brain is extremely complex, as we know, and involves a complicated interplay between functional infor- mation interacting with a structural (but not static) substrate. Brain imaging technology provides a way to sample various aspects of the brain albeit incompletely, providing a rich set of multitask and multimodal information. The field has advanced significantly in its approach to multimodal data, as there are more studies correlating, e.g., functional and structural measures. However, the vast majority of studies still ignore the joint information among two or more modalities or tasks. Such information is critical to consider as each brain imaging modality reports on a different aspect of the brain (e.g., gray matter volume, blood flow changes, white matter integrity). Multi- modal studies typically estimate features separately and combine or take an asymmetric approach (e.g., struc- ture->function). This is important as the field is still striving to understand how to diagnose and treat complex mental illness, such as schizophrenia, bipolar disorder, depression, and others, and ignoring the joint information among tasks and modalities misses a critical, but available, part of the puzzle. Combining multimodal imaging data is not easy since, among other reasons, the combination of multiple data sets consisting of thousands of voxels or timepoints yields a very high dimensional problem, requiring appropriate data reduction strategies. In the previous phase of the project, we developed new approaches to capture high-dimensional relationships among 2 or more modalities. Our work continues to strongly support the benefits of multimodal data fusion to both provide a more complete picture of brain function and structure, but also to improve our ability to study and predict the impact of complex mental illness. In this new phase of the project, we will focus on developing next generation approaches to jointly (symmetrically) estimate multimodal decomposition of matched and mis- matched data. We also focus on a novel framework called holographic fusion, allowing us to reconstruct and visualize different views of essential interactions among multimodal data from a universal model. We will evaluate this with new model-based and generative simulation frameworks. The models will be thoroughly tested by apply to a large data set which includes multiple illnesses that have overlapping symptoms, and which can sometimes be misdiagnosed and treated with the wrong medications for months or years (schizophrenia, bipolar disorder, and unipolar depression). Importantly, we also focus on early prediction. Finally, we will provide open-source tools which integrate with existing software ecosystem tools and formats and release simulation and modeling results as well as throughout the duration of the project via GitHub and the NITRIC repository, hence enabling other investigators to compare their own methods with our own as well as to apply them to a large variety of brain disorders. 35