Friday 04/15/16 1:30PM at NSH 3001

[Predrag Punosevac]

Dear Autonians,

This is a friendly remainder that this Friday's brainstorming session
speaker is Junier Oliva. The title and the abstract of his talk can be
found below. Lunch will be served as usual.

Best,
Predrag



Functional Linear Models for Brain Data

*Background.* Modern neuroimaging data has provided a much needed window
into the intricacies of the human brain. Neuroimaging techniques such as
functional magnetic resonance imaging (fMRI), magnetoencephalography
(MEG), and diffusion tensor imaging (DTI), often contain many thousands
of functional observations per subject. While some success has been had
using heuristical summary statistics of neuroimaging functional
observations to build predictive models, there is a lack of a flexible
and principled framework to build models from neuroimaging techniques.

*Aim.* Our aim is to develop a general, principled framework for
supervised learning that scales to many thousands of neuroimaging
functional observations per subject without resorting to heuristical
summary statistics. In particular, we look to regress a subject???s age
given neuroimaging data.

*Data.*Diffusion MR images were acquired from a total of 90 subjects.
The subjects ranged in age from 13 to 60 years old, and included 45
males and 45 females. The subjects had no known history of neurological
or mental disorder. We used per-voxel functional covariates of diffusion
orientation distribution functions (dODFs) in our models and real-valued
fractional anisotropy (FA) covariates as a baseline. dODFs are functions
that represent the amount of water molecules, or spins, undergoing
diffusion indifferent orientations over the S 2 sphere. FA values are
the peak of the dODF and have previously been used as a summary
statistic of dODFs for age regression. In total, 25K voxels were
considered in a shared template-space.

*Methods.* We develop and empirically test linear functional models for
neuroimaging data. We use a model that represents neuroimaging
functional observations nonparametrically using orthonormal basis
projections. The age response is modeled as a sparse additive linear
combination of inner products. We show that our model may be optimized
using a group-LASSO approach.*Results.* Our functional sparse linear
model was able to predict age with a 29.15% lower MSE than previous
heuristic-based methods. Results were found to be statistically
significant with a p-value of 0.04 using a paired t-test.

*Conclusions.* In this work we have developed and tested a principled
framework for building predictive supervised models with functional
neuroimaging data. This framework allows one to use functional
neuro-data in a statistically principled fashion without resorting to
heuristical summary statistics.