Learning Dynamic Functional Connectivity Networks from ...rahuln/pdf/2017... · Learning Dynamic...
1
Learning Dynamic Functional Connectivity Networks from Infant Magnetoencephalography Data Rahul Nadkarni ([email protected]), Nicholas J. Foti ([email protected]), Emily B. Fox ([email protected]) Goal Problem Setup Description of model Simulation Experiments Discussion Future Directions u (s) n,t = A t u (s) n,t-1 + ✏ (s) n,t ✏ (s) n,t 2 R R ⇠ N (0,Q) ROI model: • We present a method for inferring directed, dynamic functional connections between brain regions from magnetoenceph- alography (MEG) data. • Additionally, the method performs source localization and estimation of cortical surface signals informed by dynamics. • We incorporate information from multiple subjects to improve estimates of global connectivity. • We could expand upon the current method to develop a hierarchical model that shares information but also allows for inter-subject variability in learn connections. • Uncertainty in the structurals could be captured by allowing regions of interest to be learned from data rather than predefined. • Single-subject experiments demonstrate that infant structural information is noisier than adult structurals. • Multi-subject results show improved performance, highlighting the importance of sharing information across subjects even with different structural information. • Infant results motivate the need to augment the model to capture uncertainty in infant structural information. • We plan to apply the model to real infant MEG data from a music intervention study. • The sensor values , forward model , and noise covariance are subject- specific; and are known, contains known and unknown components. • The source-space noise and dynamics &:( are global and unknown, learned from data using an EM algorithm. • The entries of &:( give the strength of directed connections at all timepoints. • We are using existing structural information from 2 adult and 2 infant subjects. • The regions of interest (ROIs) are predefined, both in the left hemisphere, one posterior and one anterior. • The sensor recordings are simulated from a linear dynamical system with known, time- varying dynamics. References B. Horwitz. The elusive concept of brain connectivity. NeuroImage, 19(2):466-470, 2003. P. C. Hansen, M. L. Kringelbach, and R. Salmelin. MEG: An Introduction to Methods. Oxford University Press, 2010. Y. Yang, E. M. Aminoff, M. J. Tarr, and R. E. Kass. A state-space model of cross-region dynamic connectivity in MEG/EEG. 30 th Conference on Neural Information Processing Systems, 2016. & ,…, ( (&) , (&) (&) (&) (.) , (.) (.) (.) gradiometers magnetometers Simulated MEG data: Results: Goal: Assess the effect of structural information quality on inferring connectivity. Sensor model: y (s) n,t = C (s) u (s) n,t + ⌘ (s) n,t ⌘ (s) n,t 2 R D ⇠ N (0,R (s) )
Transcript of Learning Dynamic Functional Connectivity Networks from ...rahuln/pdf/2017... · Learning Dynamic...
Learning Dynamic Functional Connectivity Networks from Infant Magnetoencephalography DataRahul Nadkarni ([email protected]), Nicholas J. Foti ([email protected]), Emily B. Fox ([email protected])
Goal
Problem Setup
Description of model
Simulation Experiments
Discussion
Future Directions
u(s)n,t = Atu
(s)n,t�1 + ✏(s)n,t ✏(s)n,t 2 RR ⇠ N (0, Q)ROI model:
• We present a method for inferring directed,
dynamic functional connections betweenbrain regions from magnetoenceph-
alography (MEG) data.
• Additionally, the method performs source
localization and estimation of corticalsurface signals informed by dynamics.
• We incorporate information from multiple
subjects to improve estimates of globalconnectivity.
• We could expand upon the current methodto develop a hierarchical model that sharesinformation but also allows for inter-subject
variability in learn connections.
• Uncertainty in the structurals could becaptured by allowing regions of interest to
be learned from data rather thanpredefined.
• Single-subject experiments demonstratethat infant structural information is noisier
than adult structurals.
• Multi-subject results show improvedperformance, highlighting the importance
of sharing information across subjects evenwith different structural information.
• Infant results motivate the need toaugment the model to capture uncertainty
in infant structural information.
• We plan to apply the model to real infant
MEG data from a music intervention study.• The sensor values 𝑦, forward model 𝐶, and noise covariance 𝑅 are subject-
specific; 𝑦 and 𝐶 are known, 𝑅 contains known and unknown components.
• The source-space noise 𝑄 and dynamics 𝐴&:( are global and unknown, learnedfrom data using an EM algorithm.
• The entries of 𝐴&:( give the strength of directed connections at all timepoints.
• We are using existing structural informationfrom 2 adult and 2 infant subjects.
• The regions of interest(ROIs) are predefined, bothin the left hemisphere, oneposterior and one anterior.
• The sensor recordings are simulated from alinear dynamical system with known, time-varying dynamics.
ReferencesB. Horwitz. The elusive concept of brain connectivity.
NeuroImage, 19(2):466-470, 2003.
P. C. Hansen, M. L. Kringelbach, and R. Salmelin. MEG: AnIntroduction to Methods. Oxford University Press, 2010.
Y. Yang, E. M. Aminoff, M. J. Tarr, and R. E. Kass. A state-spacemodel of cross-region dynamic connectivity in MEG/EEG. 30th
Conference on Neural Information Processing Systems, 2016.
𝐴&,… , 𝐴(
𝐶(&), 𝑅(&)𝑢(&)
𝑦(&)
𝐶(.), 𝑅(.)𝑢(.)
𝑦(.)
gradiometers
magnetometers
Simulated MEG data:
Results:
Goal: Assess the effect of structural information quality on inferring connectivity.
Sensor model: y(s)n,t = C(s)u(s)n,t + ⌘(s)n,t ⌘(s)n,t 2 RD ⇠ N (0, R(s))
