multi-way modelling and analysis of high-dimensional eeg...
TRANSCRIPT
Multi-way modelling and analysis ofhigh-dimensional EEG data
Roman Rosipal
Department of Theoretical MethodsInstitute of Measurement Science, SAS
Bratislava, Slovak Republic&
Pacific Development and Technology, LLCPalo Alto, CA
FMFI UK, February 25, 2013, Bratislava
Introduction Data & Methods Results Conclusions & Beyond
Work carried out with:
Leonard J Trejo & Paul Nunez
Introduction Data & Methods Results Conclusions & Beyond
Estimation of Cognitive Status
Reward system
Executive
control
Working memory Sensation
& perception
Autonomic system
Other Processes
Internal Processes Aroused
or Overloaded
Fatigued
Resting Engaged
Working
Other States
Behavior and Performance
Biosignals
Introduction Data & Methods Results Conclusions & Beyond
Useful Definitions
Work- load
Mental Fatigue
Non- specific Factors
Engage- ment
General Cognitive
Status
Engagement: selection of a task as the focus of attention andeffort
Workload: significant commitment of attention and effort to taskOverload: task demands outstrip performance capacityMental Fatigue: desire to withdraw attention and effort from atask
Introduction Data & Methods Results Conclusions & Beyond
Useful Definitions
Work- load
Mental Fatigue
Non- specific Factors
Engage- ment
General Cognitive
Status
Engagement: selection of a task as the focus of attention andeffortWorkload: significant commitment of attention and effort to task
Overload: task demands outstrip performance capacityMental Fatigue: desire to withdraw attention and effort from atask
Introduction Data & Methods Results Conclusions & Beyond
Useful Definitions
Work- load
Mental Fatigue
Non- specific Factors
Engage- ment
General Cognitive
Status
Engagement: selection of a task as the focus of attention andeffortWorkload: significant commitment of attention and effort to taskOverload: task demands outstrip performance capacity
Mental Fatigue: desire to withdraw attention and effort from atask
Introduction Data & Methods Results Conclusions & Beyond
Useful Definitions
Work- load
Mental Fatigue
Non- specific Factors
Engage- ment
General Cognitive
Status
Engagement: selection of a task as the focus of attention andeffortWorkload: significant commitment of attention and effort to taskOverload: task demands outstrip performance capacityMental Fatigue: desire to withdraw attention and effort from atask
Introduction Data & Methods Results Conclusions & Beyond
Why to monitor cognitive status?
Critical safety, high workload, stressful, etc., environments
Introduction Data & Methods Results Conclusions & Beyond
Experiments - (A) Cognitive Workload Monitoring
Uninhabited Air Vehicle (UAV) control
Trained subjects were monitoring several UAVs as they flew apreplanned mission; processing SAR images (synthetic apertureradar), vehicle health control, etc.Different task conditions were used to control cognitive workloadlevels
Introduction Data & Methods Results Conclusions & Beyond
Experiments - (A) Cognitive Workload Monitoring
Uninhabited Air Vehicle (UAV) control
Trained subjects were monitoring several UAVs as they flew apreplanned mission; processing SAR images (synthetic apertureradar), vehicle health control, etc.
Different task conditions were used to control cognitive workloadlevels
Introduction Data & Methods Results Conclusions & Beyond
Experiments - (A) Cognitive Workload Monitoring
Uninhabited Air Vehicle (UAV) control
Trained subjects were monitoring several UAVs as they flew apreplanned mission; processing SAR images (synthetic apertureradar), vehicle health control, etc.Different task conditions were used to control cognitive workloadlevels
Introduction Data & Methods Results Conclusions & Beyond
Experiments - (B) Mental Fatigue Monitoring
Problem solving≈ 4 – 15 s
Intertrial interval 1sResponse
Problem solving≈ 4 – 15 s
3+5-7+1 <=>2??
9-6+2-4 <=>2??
Continuos performance of mental arithmetic for up to three hours
Introduction Data & Methods Results Conclusions & Beyond
Data - Electroencephalogram (EEG)
Cerebral Cortex • the outermost layers of brain • 2-4 mm thick (human)
Introduction Data & Methods Results Conclusions & Beyond
Data - EEG Sources
Introduction Data & Methods Results Conclusions & Beyond
Data - EEG Sample
Introduction Data & Methods Results Conclusions & Beyond
Spectral EEG Data Representation
Data usually segmented into epochs (e.g. 2 sec long)
Spectral representation: Estimate of the power spectrum density;that is the distribution of power per unit frequency (Thompsonmulti-taper , Welch, etc.)
Pxx (f ) = Fx (f )F ∗x (f )
where Fx (f ) is the Fourier transform of the signal x and ∗indicates the complex conjugateExample:
5 10 15 20 25 300
0.5
1
1.5
2
2.5
3
Hz
μV2 /H
z
An examle of the power spectral density estimate Subject B, electrode Cz
Introduction Data & Methods Results Conclusions & Beyond
Spectral EEG Data Representation
Data usually segmented into epochs (e.g. 2 sec long)Spectral representation: Estimate of the power spectrum density;that is the distribution of power per unit frequency (Thompsonmulti-taper , Welch, etc.)
Pxx (f ) = Fx (f )F ∗x (f )
where Fx (f ) is the Fourier transform of the signal x and ∗indicates the complex conjugate
Example:
5 10 15 20 25 300
0.5
1
1.5
2
2.5
3
Hz
μV2 /H
z
An examle of the power spectral density estimate Subject B, electrode Cz
Introduction Data & Methods Results Conclusions & Beyond
Spectral EEG Data Representation
Data usually segmented into epochs (e.g. 2 sec long)Spectral representation: Estimate of the power spectrum density;that is the distribution of power per unit frequency (Thompsonmulti-taper , Welch, etc.)
Pxx (f ) = Fx (f )F ∗x (f )
where Fx (f ) is the Fourier transform of the signal x and ∗indicates the complex conjugateExample:
5 10 15 20 25 300
0.5
1
1.5
2
2.5
3
Hz
μV2 /H
z
An examle of the power spectral density estimate Subject B, electrode Cz
Introduction Data & Methods Results Conclusions & Beyond
Coherence EEG Data Representation
Coherence representation: Cross power spectra density Pxy (f ),
Pxy (f ) = Fx (f )F ∗y (f )
or magnituted squared (coherence)
Cxy (f ) =|Pxy (f )|2
Pxx (f )Pyy (f )
Introduction Data & Methods Results Conclusions & Beyond
Data - Multi-modal Multi-Sensor
ECG - hear rate, heart rate variabilityEOG and eyes control - hEOG, vEOG movements, blinks, pupildiameterEMGSkin conductance, SCR, GCRVideotaped recordingsResponse time, Correctness of responsesSubjective responses and questionnairesetc.
Introduction Data & Methods Results Conclusions & Beyond
Data Structure
Workload
cond+ons
(e.g., trials, +me)
EEG Frequency
Electr
odes
Data matrix construction: X(I×J×K )
I - time segmentsJ - electrodes or electrode pairsK - PSD or CSD (coherences)
Introduction Data & Methods Results Conclusions & Beyond
Bilinear Unfolding
Dim 1 Dim 2 Dim 3
Dim
2 &
3
X X
X1 X2 X3 D
im 1
& 3
Dim
1 &
2
Representing all experimental factors in one dimension &observations (trials) in second dimensionContrast each dimension vs. pair of the other two
Introduction Data & Methods Results Conclusions & Beyond
Bilinear Unfolding - Modelling
Factor Analysis
Principal Component Analysis (PCA)
ijjf
F
fifij ebax +=∑
=1
∑=
=F
f 1 af
bf
0=ije
Introduction Data & Methods Results Conclusions & Beyond
Bilinear Unfolding - Regression/Classification
Partial Least Squares
Ø Data sets: X (nobjects x Nvariables)
Y (nobkjects x Mresponses) Ø Bilinear decomposition: X = TPT + E Y = UQT + F where: T,U matrices of score vectors (LV, components) P,Q matrices of loadings E,F matrices of residuals (errors) Ø Criterion: max|r|=|s|=1[cov(Xr, Ys)]2 = [cov(Xw, Yc)]2
= var(Xw)[corr(Xw, Yc)]2var(Yc) = [cov(t,u)]2
Introduction Data & Methods Results Conclusions & Beyond
NIPALS
PLS - bilinear decomposition of X and Y maximizing covariancebetween score vectors t = Xw and u = Yc
max|r|=|s|=1[cov(Xr,Ys)]2 = [cov(Xw,Yc)]2
= var(Xw)[corr(Xw,Yc)]2var(Yc)= [cov(t,u)]2
NIPALS algorithm finds the weights w,c :
1) w = XT u/(uT u) 4) c = YT t/(tT t)2) ‖w‖ → 1 5) ‖c‖ → 13) t = Xw 6) u = Yc
7) go to 1)
Introduction Data & Methods Results Conclusions & Beyond
Eigenproblem & Deflation
instead of NIPALS we can solve an eigenproblem:
w ∝ XT u ∝ XT Yc ∝ XT YYT t ∝ XT YYT Xw
XT YYT Xw = λwt = Xw
or XXT YYT t = λtu = YYT t
p = XT t/(tT t) ; q = YT u/(uT u)
sequential extraction of {ti}mi=1
X0 = Xti = Xi−1wi , Xi = Xi−1 − tipT
i = X−∑i
j=1 tjpTj
deflation schemes define different forms of PLS
Introduction Data & Methods Results Conclusions & Beyond
CCA, PLS, and PCA CR
PLS:
max|r|=|s|=1
[cov(Xr,Ys)]2 = max|r|=|s|=1
var(Xr)[corr(Xr,Ys)]2var(Ys)
CCA:max|r|=|s|=1
[corr(Xr,Ys)]2
PCA:max|r|=1
[var(Xr)]
Continuum Regression (Stone & Brooks’90)
max|r|=|s|=1
var(Xr)γ1 [corr(Xr,Ys)]2var(Ys)γ2
γ1, γ2 ∈ (0,∞)
Introduction Data & Methods Results Conclusions & Beyond
Canonical Ridge Analysis - CCA PLS
(Vinod’76)
([1− γX ]XT X + γX I)−1XT Y([1− γY ]YT Y + γY I)−1YT Xw = λw
CCA: γX = 0, γY = 0PLS: γX = 1, γY = 1Orthonormalized PLS: γX = 1, γY = 0 or γX = 0, γY = 1Ridge Regression, Regularized FDA or CCA: γX ∈ (0,1), Y ∈ R
Introduction Data & Methods Results Conclusions & Beyond
PLS Discrimination/Classification
Y =
1n1 0n1 . . . 0n1
0n2 1n2 . . . 0n2
......
. . . 1ng−1
0ng 0ng . . . 0ng
Orthonormalized PLSY = Y(YT Y)−1/2
YT Y = I
Introduction Data & Methods Results Conclusions & Beyond
Bilinear Unfolding - PLS - Classification
Introduction Data & Methods Results Conclusions & Beyond
Nonlinear (kernel) mapping: an example
Introduction Data & Methods Results Conclusions & Beyond
Example: All Degree d=2 monomials
Φ : X = R2 → F = R3
x = (x1, x2)→ Φ(x) = (x21 ,√
2x1x2, x22 )
K (x,y) = 〈Φ(x),Φ(y)〉 = (x21 ,√
2x1x2, x22 )T (y2
1 ,√
2y1y2, y22 )
= x21 y2
1 + 2x1x2y1y2 + x22 y2
2= ((x1, x2)T (y1, y2))2
= 〈x,y〉2
Problem: 〈x,y〉d ,X ⊆ RN → (N+d−1)!d!(N−1)!
N = 16× 16,d = 5→ dim. 1010
Introduction Data & Methods Results Conclusions & Beyond
Bilinear Unfolding - (Kernel) PLS - Regression
Rosipal,R & Trejo, LJ (2001). Kernel Partial Least Squares Regression inReproducing Kernel Hilbert Space. Journal of Machine Learning Research,2(Dec):97-123.
Introduction Data & Methods Results Conclusions & Beyond
Multi-way Analysis
PARAFAC
ijkkfjf
F
fifijk ecbax +=∑
=1
∑=
=F
f 1 af
bf cf
Introduction Data & Methods Results Conclusions & Beyond
PARAFAC model
The PARAFAC model with F factors: decomposition of the datamatrix X using three loading matrices, A, B, and C with elementsaif , bjf , and ckf
xijk =F∑
f=1
aif bjf ckf + εijk
The criterion:
minaif ,bjf ,ckf
= ‖xijk −F∑
f=1
aif bjf ckf‖2
Introduction Data & Methods Results Conclusions & Beyond
Multi-way PLS
Multi-way PLS (n-PLS)
X
frequency
workload condi1on
X
∑=
=F
f 1 af
cf bf
∑=
=F
f 1 vf
uf
1me
1me
max. covariance
electr
odes
EEG
Labels
af – spectral atom bf – spatial atom cf – temporal atom
vf – workload atom uf – temporal atom
Introduction Data & Methods Results Conclusions & Beyond
Mental Fatigue - PLS analysis
Black = Alert Red = Mentally Fatigued Fz Pz
Frontal Theta
Parietal Alpha
Introduction Data & Methods Results Conclusions & Beyond
Mental Fatigue - PLS analysis
Introduction Data & Methods Results Conclusions & Beyond
Mental Fatigue - PLS analysis
Robust EEG-Based Classification of Mental Fatigue 2300 (Day 1) vs. 1900 Hrs (Day 2)
40
50
60
70
80
90
100
Signal-to-noise Ratio (dB)
Test
Pro
porti
on
Corr
ect
21 Channels 100 100 99 94 79 56 50 12 Channels 98 98 96 88 65 51 50 4 Channels 87 88 90 88 76 54 50
0 -3 -6 -9 -12 -15 -18
Introduction Data & Methods Results Conclusions & Beyond
Mental Fatigue - Spectrum Analysis - PARAFAC
APECS Final Report ARO Contract No. W911NF08CO121 September 15, 2009 PDT Report No. UA01BF0636B131
Page A3-2
1200 2400 3600 4800 6000 7200 8400 9600 10800-6
-4
-2
0
2
4
6
Time(s)
Load
ing
(bas
elin
e ad
just
ed z
- sco
re)
5 10 15 20-0.2
0
0.2
0.4
0.6
0.8
1
1.2
Frequency(Hz)
Load
ing
O2 O1 OZ PZ P4 CP4 P8 C4 TP8 T8 P7 P3 CP3CPZ CZ FC4FT8 TP7 C3 FCZ FZ F4 F8 T7 FT7 FC3 F3 FP2 F7 FP1-0.3
-0.2
-0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
Electrodes
Load
ing
Subject: GSD Epoch: 2 s Overlap: 0 s Method: Welch
Figure 33. Atomic decomposition of EEG from participant GSD of the NASA-C study. EEG recordings from 30 channels were processed using PARAFAC decomposition to yield a model consisting of four atoms, each have dimensions of space (electrodes), frequency (power spectral density) and time (time on task). Graphical conventions are the same as in Figure 32. This participant performed the task for three hours, or 12 15-minute blocks. The time axis measures seconds as multiples of 2-second long EEG epochs which were not all contiguous, due to rejection of EEG segments containing movement or other artifacts. Some blocks have fewer epochs than others because the incidence of EEG artifacts increased during those blocks.
Introduction Data & Methods Results Conclusions & Beyond
Mental Fatigue - Coherence Analysis - PARAFAC
APECS Final Report ARO Contract No. W911NF08CO121 September 15, 2009 PDT Report No. UA01BF0636B131
Page A4-3
Figure 45. Coherence analyses for participant GSD. Graphing conventions are explained in Figure 44.
Introduction Data & Methods Results Conclusions & Beyond
Workload - UAV - PARAFAC
Subjects E,G,I, K (plotted subject E)
Introduction Data & Methods Results Conclusions & Beyond
Workload - UAV - PARAFAC
Subjects E - coherence
We found the similar decomposition for subjects B, G, I, K
Introduction Data & Methods Results Conclusions & Beyond
Workload - UAV - PARAFAC
Subjects E - coherence
We found the similar decomposition for subjects B, G, I, K
Introduction Data & Methods Results Conclusions & Beyond
Workload - UAV - PARAFAC
Subjects B,E,G,I, K - coherence
Introduction Data & Methods Results Conclusions & Beyond
Workload - UAV - Coherence Analysis
6 10 14 180
0.1
0.2
0.3
0.4
0.5
0.6
Frequency(Hz)
0 200 400 600 800-6
-4
-2
0
2
4
6
Time(s)
Subject: I Factor: 1 Perc.: 90
O2O1 OZ
PZ P4
CP4
P8
C4
TP8
T6
P7P3
CP3 CPZ
CZ
FC 4 T4
TP7
C3
FC Z
FZ F4F8
T5
T3 FC 3
F3
FP2
F7
FP1
Subject: I Factor: 2 Perc.: 90
O2O1 OZ
PZ P4
CP4
P8
C4
TP8
T6
P7P3
CP3 CPZ
CZ
FC 4 T4
TP7
C3
FC Z
FZ F4F8
T5
T3 FC 3
F3
FP2
F7
FP1
Subject: I Factor: 3 Perc.: 90
O2O1 OZ
PZ P4
CP4
P8
C4
TP8
T6
P7P3
CP3 CPZ
CZ
FC 4 T4
TP7
C3
FC Z
FZ F4F8
T5
T3 FC 3
F3
FP2
F7
FP1
Factor 1Factor 2Factor 3
Introduction Data & Methods Results Conclusions & Beyond
Conclusions
Results show that mental workload may be tracked by EEGcomponents isolated using PARAFAC
On UAV data set, the workload related atoms was remarkablystable in 5 out of the 6 subjects
The short-and long range coherence related atoms are morestable across the subjects, provide higher discrimination of thelow and high workload levels and seem to be less susceptible tothe movement related artifacts
We observed similarly promising and remarkable results onadditional two data sets monitoring cognitive status
Introduction Data & Methods Results Conclusions & Beyond
Future Plans
BCI, VEGA & APVV with Centre for Cognitive Science, FMFI
Resting State Networks, neurofeedback and lateralised attentionproject with Eran Zaidel lab, UCLA
Introduction Data & Methods Results Conclusions & Beyond
Thank you !!!
http://aiolos.um.savba.sk/ roman/
References:Trejo L.J., Rosipal R., Nunez P.L. Advanced Physiological Estimation of Cognitive Status (APECS). Final project report,U.S. Army Research Offfice, Research Triangle Park, NC, September 2009.Trejo L.J., Rosipal R., Nunez P.L. Advanced Physiological Estimation of Cognitive Status. The 27th Army ScienceConference, Orlando, Florida, November 29 - December 2, 2010.Rosipal, R., Trejo, L. J., Nunez, P. L. (2009). Application of Multi-way EEG Decomposition for Cognitive WorkloadMonitoring. In Proceedings of the 6th International Conference on Partial Least Squares and Related Methods, Vinzi V.E,Tenenhaus M., Guan R. (eds.), Beijing, China, pp. 145-149, 2009.
Trejo L.J., Knuth K., Prado R., Rosipal R., et al. (2007). EEG-based Estimation of Mental Fatigue: Convergent Evidence
for a Three-State Model. In Proceedings HCII 2007, Beijing, China, Springer, pp. 201-211.