A novel non-orthogonal base decoding method for fMRI neural activation
Poster No:
1960
Submission Type:
Abstract Submission
Authors:
Rongquan Zhai1
Institutions:
1Fudan University, Shanghai, Shanghai
First Author:
Introduction:
The human brain often conducts behavioral processes involving different underlying types of parallel
processing mediated by functionally distinct but spatially overlapping neural networks(Price, C.J. &
Friston, K.J.,2005). Before, human functional neuroimaging studies had difficulty revealing these
processes from underlying complex physiological signals(Zhang, Z., et al.,2017), making it difficult to
model specific processes and mechanisms of the brain(Kragel, P.A., Koban, L., Barrett, L.F. & Wager,
T.D.,2018). Recent general frameworks propose two different cognitive Processes that engage in parallel
during emotional tasks, evaluation (scaled signal value from reward to punishment) and response
readiness (contains arousal and attentional salience to aid in the response readiness process)(Fellows,
L.K.,2004;Zald, D.H. & Treadway, M.T.,2017).Evaluation and readiness signals are inevitably confounded
with each other during the emotion and the reward/punishment processing. Also, the response strength
differ from person to person which will cause the unbalanced base(the non-orthogonal bases in this
paper). In the present study, we propose a novel approach to decompose the confouded and non-orthogonal based brain activations.
processing mediated by functionally distinct but spatially overlapping neural networks(Price, C.J. &
Friston, K.J.,2005). Before, human functional neuroimaging studies had difficulty revealing these
processes from underlying complex physiological signals(Zhang, Z., et al.,2017), making it difficult to
model specific processes and mechanisms of the brain(Kragel, P.A., Koban, L., Barrett, L.F. & Wager,
T.D.,2018). Recent general frameworks propose two different cognitive Processes that engage in parallel
during emotional tasks, evaluation (scaled signal value from reward to punishment) and response
readiness (contains arousal and attentional salience to aid in the response readiness process)(Fellows,
L.K.,2004;Zald, D.H. & Treadway, M.T.,2017).Evaluation and readiness signals are inevitably confounded
with each other during the emotion and the reward/punishment processing. Also, the response strength
differ from person to person which will cause the unbalanced base(the non-orthogonal bases in this
paper). In the present study, we propose a novel approach to decompose the confouded and non-orthogonal based brain activations.
Methods:
In this study, we propose a novel method to decompose each participant's brain activations under
different circumstance(denoted as y) with a set of non-orthogonal basis (in the simulation, the base is
set to be the form like x=(x1,x2,x3))，where each vector will represent a predefined signal model,e.g.
evaluation or readiness. Specifically,'non-orthogonal' in this study means that any pairwise covariance of
vectors all not equal zero,i.e. cov(xi,xj)≠0. If cov(xi,xj)=0 (balanced base like (-1,0,1), (1,0,1)), the
regression coefficients β (i.e. the strength of signals for each individual) estimated from a multiple linear
model with all vectors were the same as those estimated univariately (of simple linear models). However,
in many tasks related to emotion or face response, the situation like the responses' strength are different
among two directions (unbalanced base like (-1,0,1), (1,0,2)) will cause cov(xi,xj)≠0. With the non-orthogonal basis, spurious correlations of signal components (i.e. β) will be introduced by related vectors
(i.e. xi are correlated), thus cause the meaningless decomposition. To overcome this difficulty, we model the signals with non-orthogonal basis with multivariate linear model (Fig1), and build a statistic (Fig1) to
separate the correlated signals and infer the independence. In Fig1.(A), we use this method to separate
simulated signals with different correlations, we can see that the statistic among varied correlations can
separate the signals well. In Fig1.(B),we use this method to identify independent signals at a fine-grained
level, in the simulations we can find the independent signals at 0.1 correlation level with sample
size=1000.
different circumstance(denoted as y) with a set of non-orthogonal basis (in the simulation, the base is
set to be the form like x=(x1,x2,x3))，where each vector will represent a predefined signal model,e.g.
evaluation or readiness. Specifically,'non-orthogonal' in this study means that any pairwise covariance of
vectors all not equal zero,i.e. cov(xi,xj)≠0. If cov(xi,xj)=0 (balanced base like (-1,0,1), (1,0,1)), the
regression coefficients β (i.e. the strength of signals for each individual) estimated from a multiple linear
model with all vectors were the same as those estimated univariately (of simple linear models). However,
in many tasks related to emotion or face response, the situation like the responses' strength are different
among two directions (unbalanced base like (-1,0,1), (1,0,2)) will cause cov(xi,xj)≠0. With the non-orthogonal basis, spurious correlations of signal components (i.e. β) will be introduced by related vectors
(i.e. xi are correlated), thus cause the meaningless decomposition. To overcome this difficulty, we model the signals with non-orthogonal basis with multivariate linear model (Fig1), and build a statistic (Fig1) to
separate the correlated signals and infer the independence. In Fig1.(A), we use this method to separate
simulated signals with different correlations, we can see that the statistic among varied correlations can
separate the signals well. In Fig1.(B),we use this method to identify independent signals at a fine-grained
level, in the simulations we can find the independent signals at 0.1 correlation level with sample
size=1000.
Results:
In real psychology experiments, we can not know the real response strength in the unbalanced basis (e.g.
(-1,0,1),(1,0,2)or(-1,0,1),(1,0,3), with our method ,we can use the model and statistic in Fig1 to find the best
base which can decompose the signals most properly. In Fig2, we set the real base is ((-1,0,1),(1,0,4)),
only the right base ((-1,0,1),(1,0,4)) has the smallest absolute t_value (two sample t test).
(-1,0,1),(1,0,2)or(-1,0,1),(1,0,3), with our method ,we can use the model and statistic in Fig1 to find the best
base which can decompose the signals most properly. In Fig2, we set the real base is ((-1,0,1),(1,0,4)),
only the right base ((-1,0,1),(1,0,4)) has the smallest absolute t_value (two sample t test).
Conclusions:
We use a generalized linear model to model the activations and build a statistic which can infer the
heterogeneity of brain signals.By implementing this method,we can identify the independent signals and
separate the signal-pairs with different correlation values.In the unbalanced base(non-orthogonal base)
settings,we can find the best base choice which can decompose the signals most properly by using this
method,it will help to design the base in the experiments with emotional tasks.In the future work,we can
extending our method to higher dimensions and whole brain(Berridge, K. C. (2019);Kauschke,C., Bahn,D.,
Vesker, M., & Schwarzer,G. (2019)).
heterogeneity of brain signals.By implementing this method,we can identify the independent signals and
separate the signal-pairs with different correlation values.In the unbalanced base(non-orthogonal base)
settings,we can find the best base choice which can decompose the signals most properly by using this
method,it will help to design the base in the experiments with emotional tasks.In the future work,we can
extending our method to higher dimensions and whole brain(Berridge, K. C. (2019);Kauschke,C., Bahn,D.,
Vesker, M., & Schwarzer,G. (2019)).
Emotion, Motivation and Social Neuroscience:
Emotional Learning
Emotional Perception
Modeling and Analysis Methods:
Activation (eg. BOLD task-fMRI) 2
Multivariate Approaches 1
Keywords:
Computational Neuroscience
fMRI CONTRAST MECHANISMS
FUNCTIONAL MRI
Modeling
1|2Indicates the priority used for review
·Fig 1.Model overview and signal seperation, independent signal identification.(A)Our method can seperate signals with different correlations.(B)Our method can identify independent signal at 0.1 level.
··Fig 2.Searching the best base.(A1-A5)Correlation values distribution with different base decoping. (B)Absolute t_values with different base decoping,correct base's absolute t_value is the smallest.
Provide references using author date format
structure and function. Cogn Neuropsychol 22, 262-275 .
2.Zhang, Z., et al.(2017) Distributed neural representation of saliency controlled value and category during
anticipation of rewards and punishments. Nat Commun 8, 1907 .
3.Kragel, P.A., Koban, L., Barrett, L.F. & Wager, T.D. (2018) Representation, Pattern Information, and
Brain Signatures: From Neurons to Neuroimaging. Neuron 99, 257-273 .
4.Fellows, L.K. (2004)The cognitive neuroscience of human decision making: a review and conceptual
framework. Behav Cogn Neurosci Rev 3, 159-172 .
5.Zald, D.H. & Treadway, M.T. (2017)Reward Processing, Neuroeconomics, and Psychopathology. Annual
Review of Clinical Psychology 13, 471-495.
6.Berridge, K. C. (2019). Affective valence in the brain: modules or modes?. Nature Reviews
Neuroscience, 20(4), 225-234.
7.Kauschke, C., Bahn, D., Vesker, M., & Schwarzer, G. (2019). The role of emotional valence for the
processing of facial and verbal stimuli—positivity or negativity bias?. Frontiers in psychology, 10, 1654.