Paul Thomas is working towards his MD/PhD in Bioengineering. Paul's research project (Olusola Ajilore, Advisor) research project aims to address the needs proposed by the RDoC initiative by employing a data-driven approach on a recently collected dataset from an RDoC study, “Negative Valence Brain Targets and Predictors of Anxiety and Depression Treatment” (PI: K. Luan Phan, MD, NIMH R01MH101497).

Paul writes:

This dataset is from a transdiagnostic sample of participants with any type of IP. Each participant was screened to confirm diagnosis, and then randomized to either CBT or SSRI base therapy. A battery of measurements, including subjective rating scales, and multiple neuroimaging scans (detailed in
Table 1.) were taken before and after 12 weeks of treatment. Healthy controls were also subjected to the same measurements and both timepoints. From the neuroimaging scans (resting state fMRI, EEG, and diffusion tensor imaging-MRI), a connectome, which is either a functional (fMRI and EEG) or structural (DTI) representation of links between different brain regions, will be generated for each individual. Each connectome can then be represented by a graph where each node is a brain region of interest (ROI) and the weights of the edges connecting the nodes to one another is defined by the strength of correlation or structural link between each node. This method of encoding the connectome allows the network properties of each graph to be
compared between subjects of the study. In this project, we investigate subject connectomic networks using each of two potential paradigms. First, we compute graph theory based network metrics for each subjects whole connectome, as well as for each node, or ROI, in each connectome. We then compare metrics between groups and over time to characterize the relationship between relatively manifest network properties and disease state. Second, we use a more data driven machine learning approach to study the classification of subjects via their connectomic properties. In this project, we use a tensor-factorization methods developed previous by our group, called tensor-based brain network embedding (t-BNE). A significant advantage of t-BNE
compared to other classifiers, is that it allows for the input of a heterogeneous dataset. This means that subjects can be classified by their connectivity matrices in addition to various auxiliary information (e.g., subjective mood rating, demographic information). Through this method, we are able to study the relatively latent factors that may be helpful for distinguishing subjects based on disorder, or response or a specific treatment modality.