Scribd
Upload
39 views36 pages

Presentation EiDA

EiDA: A lossless approach for the dynamic analysis of connectivity patterns in signals; application to resting state fMRI of a model of ageing

Uploaded by

dagush
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
39 views36 pages

Presentation EiDA

EiDA: A lossless approach for the dynamic analysis of connectivity patterns in signals; application to resting state fMRI of a model of ageing

Uploaded by

dagush
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 36

CNS, JC talks

EiDA
Gustavo Patow
ViRVIG-UdG, CNS-UPF

[email protected]

10/oct/2023 1
Eider’s challenge!

JC - Oct 10th, 2023 2


TL;DR
•EiDA
• Same as LEiDA, but without L
• Really!

JC - Oct 10th, 2023 3


CNS, JC talks

EiDA
Thanks!!!
Did I win?

4
A preprint… https://www.biorxiv.org/content/10.1101/2023.02.27.529688

DA
Ei
ri is /
l te
/a
om
: b . c
e
d hu
C o i t
g
ith s://
W tp
JC - Oct 10th, 2023
ht 5
Leading Eignevector Dynamics
Analysis

State1 Probabilistic StateN Probabilistic


Metastable Substate Metastable Substate

Let’s see each stage in more detail…

JC - Oct 10th, 2023 6


LEiDA Computations (I)
• BOLD phase synchronisation
• Compute BOLD phase
• Hilbert: xa = a(t)ei𝜙(t)

JC - Oct 10th, 2023 7


LEiDA Computations (II)
• Phase dFC and Leading Eigenvectors
• Calculate phase synchrony:
dFCn,p(t) = cos(𝜙n(t) - 𝜙p(t))

• Called iPLn,p(t) in the paper…

JC - Oct 10th, 2023 8


LEiDA Computations (III)
• Clustering and probabilistic metastable substate (PMS)

State1 Probabilistic StateN Probabilistic


Metastable Substate Metastable Substate

JC - Oct 10th, 2023 9


EiDA proposed method (old part)

Hilbert transform
x(t) A(t) exp(i 𝜃(t))
x(t) + i 𝓗{x(t)}
angle
JC - Oct 10th, 2023 10
EiDA proposed method (new part)

Analysis
JC - Oct 10th, 2023 11
The iPL (dFC) Matrix
• Remember:
iPLn,p(t) = cos(𝜙n(t) - 𝜙p(t))
• But cos(a-b) = cos(a) cos(b) + sin(a) sin(b)
• Thus iPLn,p(t) = cos(𝜙n(t)) cos(𝜙p(t)) + sin(𝜙n(t)) sin(𝜙p(t))
• We can define
c = (cos(𝜙1(t)), cos(𝜙2(t)), …, cos(𝜙N(t)))T
s = (sin(𝜙1(t)), sin(𝜙2(t)), …, sin(𝜙N(t)))T
• And rewrite
iPL(t) = c cT + s sT

JC - Oct 10th, 2023 12


The iPL (dFC) Matrix
• If
iPL(t) = c cT + s sT
• Then the iPL(t) is a rank 2 matrix!
• Symmetric
• Positive semidefinite
• Will have only 2 non-null e-values 𝜆1 and 𝜆2, with their
e-vectors
• Even more, Tr(iPL) = N = 𝜆1 + 𝜆2

JC - Oct 10th, 2023 13


Fu
ll d
p a em
The iPL (dFC) Matrix p e on
r a str
pp ati
e n on
dix @

• Analytic formulation of
e-vectors and e-values!

• The e-vectors will be of the form


Ac+Bs

• And the e-values will be

JC - Oct 10th, 2023 14


From the discussion @ JC
• By construction 𝜆1 > 𝜆2
• Because B1 takes the + sign and B2 the – sign in the square root
(previous slide)
• Thus, 1st e-vector will always be the largest
• 𝜆2 ∈ [1, 22] and 𝜆1 ∈ [23, 44] (remember, N = 𝜆1 + 𝜆2)
• Assuming arrays stat at 1 (Matlab)
• They will always be sorted!

JC - Oct 10th, 2023 15


Cohort…

N = 44 RoIs
Nsub = 48
fMRI

JC - Oct 10th, 2023 16


Illustration
λ1 = N, λ2 = 0 λ1 = λ2 = N/2

λ1 = 39 (/44)
LEiDA

Exemplars
from the rat
λ1 = 23 (/44)
data EiDA

Cannot be discarded! 17
JC - Oct 10th, 2023
Results from the
static analysis
<FC>N

<iPL>t

FC index = 𝛴ij FCij2


!
iii) Best Ordinary Least Squares fit
18
JC - Oct 10th, 2023
Two new analysis approaches
•Discrete EiDA
•Continous EiDA

JC - Oct 10th, 2023 19


Discrete EiDA K-Means algorithm

Main Points:
• “Discrete” means Clustering!
• 2 evectors == 1 datapoint
• Rebuild iPL, take upper triang matrix
• Standard k-means on this
• Cosine “distance”

JC - Oct 10th, 2023 20


Discrete EiDA K-Means algorithm
Full details…

21
JC - Oct 10th, 2023
Some definitions…
• Dwell time
avg duration of a mode (cluster)
• Fractional occurrence
average amount of time where the iPL ∈ mode
• Spectral Metastability (N = λ1 + λ2)
metaspec = std(λ1) = std(λ2)

JC - Oct 10th, 2023 22


Results for the
Discrete EiDA

3 cluster centroids

(sum of dist2 of each datapoint w/ their centroids)

JC - Oct 10th, 2023 23


Continous EiDA
• Clusters may be blurry, mixed up…
• Poor convergence
• Lack of spatial heterogeneity
• Continous exploration of Phase Locking configs
• Plot evolution of both evectors over time => Continous EiDA
• Use kinematic speed-displacement (KSD) plot

JC - Oct 10th, 2023 24


Continous EiDA
• kinematic speed-displacement (KSD)
• Position: spectral radius = λ1
• Speed:
s(t) = 1 – |R(iPL(t), iPL(t-1))|
with R the Pearson Corr Coef
• Alternative definition: sepparately for evectors
• Position: λ1 and λ2
• Speed as in s(t), but correlating evectors

JC - Oct 10th, 2023 25


Results for the continuous EiDA
For one representative rat…
+ = center of mass of the distr

Broader to more compact


=> Connectivity patterns from more structured (less speed,
higher evalues) to more random…

JC - Oct 10th, 2023 26


Some more definitions…
• Spectral radius (of the iPL matrix)
λ1 (Closed form norm of the iPL matrix)
• Information Complexity
Lempel-Zip-Welch (LZW) algorithm

Used in gzip and pkzip!


• The code saves data as a
.mat file and gzip’s it!
• Directly applied over the
whole BOLD signal!
JC - Oct 10th, 2023 27
Results for the continuous EiDA
Decrease over time! Inverse relation!

JC - Oct 10th, 2023 28


Results for the continuous EiDA

Irreducibility Index: proportion of


time in which λ1 is lower than X (%)

Careful wording here: fraction of time where reduction to


1st evector would fail in keeping this information (W.R.T.
LEiDA) !
JC - Oct 10th, 2023 29
Results for the continuous EiDA

More compact
Faster with ageing! exploration space
with ageing!
JC - Oct 10th, 2023 30
Results for the continuous EiDA
Quadratic fit

All N=48 rats


JC - Oct 10th, 2023 31
Position-speed plots of the two
separate eigenvectors
same representative rat as before

JC - Oct 10th, 2023 32


EiDA transformation from the raw data
to the evolution of the two e-vectors
same representative rat

Average
reconfiguration speed

Higher instantaneous phase locking in the raw signals Lack of instantaneous phase locking
(more ordered and compact evolution, first eigenvector (more random, less structured and
contains most of the information) equally informative eigenvectors)
JC - Oct 10th, 2023 33
Contributions
• A full characterization of the iPL matrix and its properties.
• An analytical and therefore ultra-fast way of decomposing it.
• A new algorithm (discrete EiDA) for the clustering of brain
states
• Ease RAM requirements by orders of magnitude
• Discrete EiDA solves the "flipping sign" problem that was observed in
LEiDA
• A 2-d embedding of dFC and two new measures for the
continuous analysis of the connectivity patterns.

JC - Oct 10th, 2023 34


Conclusion
• Explored the implications of using phase relationships and
the iPL
• Reconfiguration dynamics in addition to state dynamics
• Lossless decomposition of the iPL
• Complete information is encoded in two eigenvectors and one
eigenvalue
• novel measures to further characterize the dynamic evolution of the
spontaneous activity in the resting-state brain.
• Application: rs-fMRI data from a study of ageing in rats
• EiDA provides an analytically based framework for
dimensionality reduction and dynamical analysis
• Discrete and continuous versions!

JC - Oct 10th, 2023 35


CNS, JC talks

EiDA
Thanks!!!
Now yes, Questions?

36

You might also like