Methods: Reproducible & Open Science
Fields: Health Science, Statistics

Project Lead: Nick Foti, WRF Postdoctoral Fellow in Neuroengineering and Data Science, Department of Statistics
Advisors: Emily Fox, Amazon Professor of Machine Learning, Department of Statistics, and Adrian KC Lee, Associate Professor, Department of Speech and Hearing Sciences and Institute for Learning and Brain Sciences

Overview

A depiction of Gaussian dataMany cognitive disorders, such as autism and (central) auditory processing disorder (C)APD are thought to arise due to deficiencies in the underlying communication network that the brain implements to exhibit specific cognitive behaviors. Inferring these communication networks from non-invasive neuroimaging data in both healthy and clinical populations would go a long way towards understanding how cognitive processes arise, and in developing therapies and assistive devices for individuals affected by these disorders.

For instance, individuals with (C)APD report that maintaining conversations in loud,crowded venues to be exhausting due to the cognitive load required to attend to different speakers. Understanding the functional connectivity underlying auditory attention could allow the development of smart hearing aids to alleviate this issue for individuals with (C)APD.

The goal of this project is to developing statistical and computational
methodology in order to infer interactions between high-dimensional time series
data arising from magneto-encephalography (MEG) recordings of subjects
performing auditory attention tasks. The inferred connectivity structures can
then potentially be used to inform future hypothesis generation and
experimental design to understand auditory attention.

We have developed a method to learn Gaussian graphical models (GGMs) that
encode conditional independencies between cortical signals, accounting for both
the dynamics of the process as well as unobserved processes that could
otherwise lead to spurious connections. In order to analyze the large amount of
MEG data that has been acquired, we have formulated the model as a convex
optimization problem and have developed efficient first- and second-order
algorithms to solve it.