Date of Award
Doctor of Philosophy (PhD)
School of Mathematical and Statistical Sciences
D. Andrew Brown
This dissertation investigates the functional graphical models that infer the functional connectivity based on neuroimaging data, which is noisy, high dimensional and has limited samples. The dissertation provides two recipes to infer the functional graphical model: 1) a fully Bayesian framework 2) an end-to-end deep model.
We first propose a fully Bayesian regularization scheme to estimate functional graphical models. We consider a direct Bayesian analog of the functional graphical lasso proposed by Qiao et al. (2019).. We then propose a regularization strategy via the graphical horseshoe. We compare both Bayesian approaches to the frequentist functional graphical lasso, and compare the Bayesian functional graphical lasso to the functional graphical horseshoe. We applied the proposed methods with electroencephalography (EEG) data and diffusion tensor imaging (DTI) data. We find that the Bayesian methods tend to outperform the standard functional graphical lasso, and that the functional graphical horseshoe performs best overall, a procedure for which there is no direct frequentist analog.
Then we consider a deep neural network architecture to estimate functional graphical models, by combining two simple off-the-shelf algorithms: adaptive functional principal components analysis (FPCA) Yao et al., 2021a) and convolutional graph estimator (Belilovsky et al., 2016). We train our proposed model with synthetic data which emulate the real world observations and prior knowledge. Based on synthetic data generation process, our model convert an inference problem as a supervised learning problem. Compared with other framework, our proposed deep model which offers a general recipe to infer the functional graphical model based on data-driven approach, take the raw functional dataset as input and avoid deriving sophisticated closed-form. Through simulation studies, we find that our deep functional graph model trained on synthetic data generalizes well and outperform other popular baselines marginally. In addition, we apply deep functional graphical model in the real world EEG data, and our proposed model discover meaningful brain connectivity.
Finally, we are interested in estimating casual graph with functional input. In order to process functional covariates in causal estimation, we leverage the similar strategy as our deep functional graphical model. We extend popular deep causal models to infer causal effects with functional confoundings within the potential outcomes framework. Our method is simple yet effective, where we validate our proposed architecture in variety of simulation settings. Our work offers an alternative way to do causal inference with functional data.
Niu, Jiajing, "Learning Graphical Models of Multivariate Functional Data with Applications to Neuroimaging" (2022). All Dissertations. 3239.
Author ORCID Identifier