a state-space model estimation of eeg signals using subspace...
TRANSCRIPT
A State-space model estimation of EEG signals using subspace identification
Satayu Chunnawong ID 5730569821Advisor: Assist. Prof. Jitkomut Songsiri
Department of Electrical EngineeringChulalongkorn University
2102490 Project Proposal Year 2017
1
Outline
Introduction Background PreliminaryResults
Methodology ProjectOverview
2
Introduction
Source : C. Pawela and B. Biswal, “Brain Connectivity : A New Journal Emerges,” Brain Connectivity, vol. 1, no. 2, 2011.
3
Learning Causality
Random Vectors
Cross Covariance
Partial Correlation
Conditional Independence
Time Series
DCM
SEM
Granger causality
AR model
State space model
Cross Covariance Function
Learning causality applied on brain connectivity
4
Pre-Processing
State-space model estimation and
selection
GC test on state-space model
SS Model(𝒜𝒜 ,𝐶𝐶,𝑊𝑊,𝑉𝑉)
Methodology
Methodology
EEG time series 𝑦𝑦(𝑡𝑡)
�𝑦𝑦(𝑡𝑡)
Outcome: Schemes for estimating state space models that infer GC
EEG signals of healthy subject
EEG signals of subject with epilepsy
GC test GC test
Causalitypattern
Outcome: Comparison results of brain connectivity
Project Scheme andExpected outcomes
Background : EEG Models
6
𝑥𝑥 𝑡𝑡 = �𝑘𝑘=1
𝑝𝑝
𝐴𝐴𝑘𝑘𝑥𝑥 𝑡𝑡 − 𝑘𝑘 + 𝑢𝑢(𝑡𝑡)
y 𝑡𝑡 = 𝐿𝐿𝑥𝑥 𝑡𝑡 + 𝑣𝑣(𝑡𝑡)
𝑥𝑥 : Sources (not measured)y : Measurement
There are many mathematical models that describe EEG signals. One can be generally described by linear Autoregressive (AR) model.
7
In this project, we focus on only two Granger causality tests :
GC test for AR model
𝐴𝐴𝑘𝑘 𝑖𝑖𝑖𝑖 = 0
𝑦𝑦 𝑡𝑡 = �𝑘𝑘=1
𝑝𝑝
𝐴𝐴𝑘𝑘𝑦𝑦 𝑡𝑡 − 𝑘𝑘 + 𝑢𝑢(𝑡𝑡)
According to Pruttiakaravanich (2016), If 𝑦𝑦𝑖𝑖 does not cause 𝑦𝑦𝑖𝑖 , it can be shown that
8
GC test for state space model (Seth, 2015)
ℱ𝑦𝑦𝑗𝑗→𝑦𝑦𝑖𝑖| 𝑎𝑎𝑎𝑎𝑎𝑎 𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜 𝑦𝑦 = logΣii𝑅𝑅
Σii
In general, prediction error of 𝑦𝑦𝑖𝑖 in reduced model is bigger prediction error of 𝑦𝑦𝑖𝑖 in full model. If Σii𝑅𝑅 = Σii , it means 𝑦𝑦𝑖𝑖 does not cause 𝑦𝑦𝑖𝑖
Let Σ, Σ𝑅𝑅 as prediction error covariance of full model and reduced model, respectively.
𝑧𝑧 𝑡𝑡 + 1 = 𝒜𝒜𝑧𝑧 𝑡𝑡 + 𝑤𝑤 𝑡𝑡𝑦𝑦 𝑡𝑡 = 𝐶𝐶𝑧𝑧 𝑡𝑡 + 𝑣𝑣 𝑡𝑡
𝑦𝑦𝑅𝑅 𝑡𝑡 = 𝐶𝐶𝑅𝑅𝑧𝑧 𝑡𝑡 + 𝑣𝑣(𝑡𝑡)(Full model)
(Reduced model)
To remove 𝑦𝑦𝑖𝑖 is to remove 𝑗𝑗𝑜𝑜𝑜 column of 𝐶𝐶
9
MethodologyStochastic Subspace Identification (Overschee and De Moor, 1996)
10
MethodologyGranger causality onState space model
State space model (full model)
𝑧𝑧 𝑡𝑡 + 1 = 𝒜𝒜𝑧𝑧 𝑡𝑡 + 𝑤𝑤 𝑡𝑡𝑦𝑦 𝑡𝑡 = 𝐶𝐶𝑧𝑧 𝑡𝑡 + 𝑣𝑣(𝑡𝑡)
𝑧𝑧 𝑡𝑡 + 1 = 𝒜𝒜𝑧𝑧 𝑡𝑡 + 𝑤𝑤 𝑡𝑡𝑦𝑦 𝑡𝑡 = 𝐶𝐶𝑅𝑅𝑧𝑧 𝑡𝑡 + 𝑣𝑣(𝑡𝑡)
State space model (reduced model)Σ: estimation error covariance of full model
ΣR : estimation error covariance of reduced model
11
MethodologyGranger causality onState space model
We find estimation error covariance : Σ = 𝑐𝑐𝑐𝑐𝑣𝑣 𝑧𝑧 − �̂�𝑧 for full model and Σ𝑅𝑅for reduced model.
�̂�𝑧𝑜𝑜+1|𝑜𝑜 = 𝒜𝒜�̂�𝑧𝑜𝑜|𝑜𝑜−1 + Σ𝑜𝑜|𝑜𝑜−1𝐶𝐶𝑇𝑇 𝐶𝐶Σ𝑜𝑜|𝑜𝑜−1𝐶𝐶𝑇𝑇 + 𝑅𝑅 −1 𝑦𝑦𝑜𝑜 − 𝐶𝐶�̂�𝑧𝑜𝑜|𝑜𝑜−1= 𝒜𝒜�̂�𝑧𝑜𝑜|𝑜𝑜−1 + 𝐾𝐾(𝑦𝑦𝑜𝑜 − �𝑦𝑦𝑜𝑜|𝑜𝑜−1)
Where �̂�𝑧 is an optimal estimation of 𝑧𝑧 which observed by using Kalman filter
12
MethodologyGranger causality onState space model
By assumption, we can solve steady stateKalman filter which satisfies DARE
Σ = 𝒜𝒜Σ𝒜𝒜𝑇𝑇 + 𝑊𝑊 −𝒜𝒜Σ𝐶𝐶𝑇𝑇 𝐶𝐶Σ𝐶𝐶𝑇𝑇 −1𝐶𝐶Σ𝒜𝒜𝑇𝑇
Then, Seth (2015) suggest to determine the time-domain Granger causality shown as
ℱ𝑥𝑥𝑗𝑗→𝑥𝑥𝑖𝑖| 𝑎𝑎𝑎𝑎𝑎𝑎 𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜 𝑥𝑥 = logΣ𝑖𝑖𝑖𝑖𝑅𝑅
Σii13
Preliminary ResultsThere are two experiments in this semester
A
Equivalence of GC test on AR and state space model
GC test on estimated state space model
Hypothesis: If ground truth is AR, the result of GC test on state space model is the same result as GC
test on AR model
Hypothesis: If EEG signals are generated based on AR model, the result of GC test on estimated state
space model is the same result as GC test on AR model14
Preliminary ResultsEquivalence of GC test
A
15
Preliminary ResultsA
We format state space model from ground truth AR model
𝑧𝑧 𝑡𝑡 + 1 = 𝒜𝒜𝑧𝑧 𝑡𝑡 + 𝑤𝑤 𝑡𝑡𝑦𝑦 𝑡𝑡 = 𝐶𝐶𝑧𝑧 𝑡𝑡 + 𝜖𝜖(𝑡𝑡)
𝒜𝒜 =
𝐴𝐴1 𝐴𝐴2 ⋯ 𝐴𝐴𝑝𝑝𝐼𝐼 0 … 0⋮ ⋱ ⋱ ⋮0 … 𝐼𝐼 0
𝐶𝐶 =𝐿𝐿𝑇𝑇0⋮0
𝑇𝑇
𝜖𝜖 𝑡𝑡 =
𝑣𝑣 𝑡𝑡0⋮0
𝑤𝑤 𝑡𝑡 =
𝑢𝑢 𝑡𝑡0⋮0
16
Preliminary Results>> F
F =
0.5207 0.0016 0.00000.2412 0.4073 0.14170.0000 0.0282 0.1385
>> F_r
F_r =
0.5207 0.0016 0.00000.2412 0.4073 0.14170.0000 0.0282 0.1385
GC test on AR model
GC test on state space model
17
Preliminary ResultsGC test on estimated state space model
18
Preliminary Results>> F
F =
0.5207 0.0016 0.00000.2412 0.4073 0.14170.0000 0.0282 0.1385
>> F_ss
F_ss =
0.2162 0.2033 0.68430.2532 3.1890 0.75490.2575 0.9480 1.4804
GC test on AR model
GC test on estimated state space model
19
20
Pre-ProcessingState-space model estimation and
selection
GC test on state-space model
SS Model(𝒜𝒜 ,𝐶𝐶,𝑊𝑊,𝑉𝑉)
EEG time series 𝑦𝑦(𝑡𝑡)
�𝑦𝑦(𝑡𝑡)
Outcome: Schemes for estimating state space models that infer GC
EEG signals of healthy subject
EEG signals of subject with epilepsy
GC test GC test
Causalitypattern
Outcome: Comparison results of brain connectivity
100% finished
70% finished
Q&A
22
Reference• L. Barnett and A. K. Seth, “Granger causality for state space models,” Physical
Review E, vol. 91, no. 4, 2015.• A. Pruttiakaravanich and J. Songsiri, “A review on dependence measures in exploring
brain networks from fMRI data,” Engineering Journal, vol. 20, no. 3, pp. 207–233, 2016
• P. van Overschee and B. de Moor, Subspace Identification for Linear Systems. Hoboken, New Jersey: Wiley, 1996
23
Backup𝐶𝐶𝒜𝒜𝑐𝑐
𝑘𝑘𝐾𝐾 coefficeints
We showed that the coefficient from GC test on AR model have same structure to the coefficient from GC test on state space based on ground truth AR model
Let 𝐾𝐾 as gain solved from steady state Kalman filter
𝐾𝐾 = 𝒜𝒜Σ𝐶𝐶𝑇𝑇 𝐶𝐶Σ𝐶𝐶𝑇𝑇 + 𝑉𝑉 −1
= 𝒜𝒜 Σ11𝑇𝑇 Σ12𝑇𝑇 … Σ1𝑝𝑝𝑇𝑇𝑇𝑇Σ11−1
Because the solution of DARE remains only Σ11𝐾𝐾 = 𝒜𝒜 𝐼𝐼 0 0 0 𝑇𝑇
= 𝐴𝐴1𝑇𝑇 𝐼𝐼 0 0 𝑇𝑇
24
Backup𝐶𝐶𝒜𝒜𝑐𝑐
𝑘𝑘𝐾𝐾 coefficeints
Given state observer gain 𝒜𝒜𝑐𝑐 = 𝒜𝒜 −𝐾𝐾𝐶𝐶 , we have
𝒜𝒜𝑐𝑐 =
0 𝐴𝐴2 ⋯ 𝐴𝐴𝑝𝑝0 0 … 0⋮ ⋱ ⋱ ⋮0 … 𝐼𝐼 0
Then, multiply by 𝐶𝐶 on the left hand side and 𝐾𝐾 on the right hand side
𝐶𝐶𝒜𝒜𝑐𝑐𝑘𝑘𝐾𝐾 = 𝐼𝐼 0 ⋯ 0
0 𝐴𝐴2 ⋯ 𝐴𝐴𝑝𝑝0 0 … 0⋮ ⋱ ⋱ ⋮0 … 𝐼𝐼 0
k
𝐴𝐴1𝑇𝑇 𝐼𝐼 0 0 𝑇𝑇
25
Backup𝐶𝐶𝒜𝒜𝑐𝑐
𝑘𝑘𝐾𝐾 coefficeints
𝐶𝐶𝒜𝒜𝑐𝑐𝑘𝑘𝐾𝐾 = 𝐼𝐼 0 ⋯ 0
0 𝐴𝐴2 ⋯ 𝐴𝐴𝑝𝑝0 0 … 0⋮ ⋱ ⋱ ⋮0 … 𝐼𝐼 0
k
𝐴𝐴1𝑇𝑇 𝐼𝐼 0 0 𝑇𝑇
The result showed that
When 𝑘𝑘 = 0 𝐶𝐶𝐾𝐾 = 𝐴𝐴1When 𝑘𝑘 = 1 𝐶𝐶𝒜𝒜𝑐𝑐𝐾𝐾 = 𝐴𝐴2When 𝑘𝑘 = 2 𝐶𝐶𝒜𝒜𝑐𝑐
2𝐾𝐾 = 𝐴𝐴3⋮
When 𝑘𝑘 = 𝑝𝑝 − 1 𝐶𝐶𝒜𝒜𝑐𝑐𝑝𝑝−1𝐾𝐾 = 𝐴𝐴𝑝𝑝
26