Article
https://doi.org/10.1038/s41467-023-43747-5
Neural substrates of parallel devaluation-
sensitive and devaluation-insensitive
Pavlovian learning in humans
Eva R. Pool
1,2
,WolfgangM.Pauli
2,3
, Logan Cross
4,5
& John P. O
’
Doherty
2,3
We aim to differentiate the brain regions involved in the learning and encoding
of Pavlovian associations sensitive to changes in outcome value from those
that are not sensitive to such changes
by combining a learning task with out-
come devaluation, eye-tracking, and functional magnetic resonance imaging
in humans. Contrary to theor
etical expectation, voxels
correlating with reward
prediction errors in the ventral striat
um and subgenual cingulate appear to be
sensitive to devaluation. Moreover, re
gions encoding state
prediction errors
appear to be devaluation insensitive. We can also distinguish regions encoding
predictions about outcome taste iden
tity from predictions about expected
spatial location. Regions encoding p
redictions about taste identity seem
devaluation sensitive while those en
coding predictions about an outcome
’
s
spatial location seem devaluation insensitive. These
fi
ndings suggest the
existence of multiple and distinct associative mechanisms in the brain and help
identify putative neural correlates fo
r the parallel expression of both deva-
luation sensitive and insensitive conditioned behaviors.
Pavlovian learning is one of the simplest and most fundamental forms
of learning, whereby an initially neutral stimulus (conditioned stimulus
or CS; e.g., a metronome sound) acquires or changes value by being
associated with an affectively signi
fi
cant outcome (e.g., food)
1
–
3
.This
form of associative learning exerts a profound in
fl
uence on behavior
4
,
5
,
cognition
6
,
7
, and mental health
8
–
10
. Despite being extensively studied
across animals and humans, the neurocomputational mechanisms
involved in Pavlovian learning appear to be more elaborate than pre-
viously conceived
11
–
16
.
Value learning signals during Pavlovian conditioning have been
extensively characterized in the brain. Reward prediction errors
–
a
learning mechanism through which the CS becomes endowed with an
outcome
’
s affective value
17
–
20
–
have been shown to correlate with
dopaminergic activity in the midbrain
21
,aswellasbloodoxygenation
level dependent (BOLD) responses in the ventral striatum
22
and
midbrain
23
. Moreover, the acquisition of affective conditioned
responses appears to involve frontomedial structures such as the
ventromedial prefrontal cortex (vmPFC)
24
and subgenual anterior
cingulate cortex (sgACC)
25
,
26
. Lesion studies in monkeys
26
and
humans
24
suggest that these structures are critical for a CS to trigger
affective conditioned responses, re
fl
ected either in pupil dilation
26
or
in skin conductance
24
.
In the last decade, a growing number of studies have found evi-
dence for other kinds of learning signals. Speci
fi
cally, neural signals
associated with model-based representations, or cognitive maps, have
been identi
fi
ed during Pavlovian learning
5
,
11
,
13
,
27
–
29
.Akeylearningsignal
suggested to be involved in the building of a cognitive map is the state
prediction error. This prediction error quanti
fi
es how unexpected a
particular perceptual state is given the previous state, independently
of its affective value and is implicated in the acquisition of a
state
–
space transition model. State prediction errors have been
reported in the lateral prefrontal cortex (PFC)
30
, lateral orbitofrontal
cortex (OFC), anterior insula, superior frontal gyrus (SFG)
31
,
32
,andthe
intraparietal cortex
5
,
30
. It has been shown that reward prediction errors
Received: 24 February 2023
Accepted: 17 November 2023
Check for updates
1
Swiss Center for Affective Sciences, University of Geneva, Geneva, Switzerland.
2
Division of Humanities and Social Sciences, California Institute of Tech-
nology, Pasadena, CA, USA.
3
Computation and Neural Systems Program, California Institute of Technology, Pasadena, CA, USA.
4
Division of Biology and
Biological Engineering, California Institute of Technology, Pasadena, CA, USA.
5
Department of Computer Science, Stanford University, Palo Alto, CA, USA.
e-mail:
eva.pool@unige.ch
Nature Communications
| (2023) 14:8057
1
1234567890():,;
1234567890():,;
in the ventral striatum exist alongside state prediction errors in other
structures such as the lateral OFC
32
.TheOFChasbeenimplicatedin
the representation of states and cognitive maps in a large corpus of
studies
33
–
37
and increasing evidence implicates OFC model-based
representations in Pavlovian learning, including representations of
outcome identity, an outcome
’
s sensory features, sensory feature
changes, and the acquisition of stimulus
–
stimulus associations
underlying the construction of cognitive maps
11
,
27
–
29
. Interestingly,
model-based learning signals during Pavlovian conditioning have also
been found in brain regions typically involved in value learning such as
the striatum
11
, amygdala
4
, and the dopaminergic midbrain
–
with
activity coding for the updating of expectations about perceptual
attributes of the outcome
28
,
38
,
39
.
The co-existence of distinct neural signals related to an outcome
’
s
affective value on one hand and the perceptual properties of an out-
come on the other, are broadly consistent with theoretical models
postulating that Pavlovian learning is not a unitary process, but rather
involves several parallel associations between the CS and multiple
attributes of the outcome
40
,
41
. There has been a long-standing con-
ceptualization of multiple and parallel conditioned responses to a
given CS
42
–
44
, but only recently have these classes of behavioral
responses and their underlying neural learning signals been investi-
gated in humans
16
,
45
. Strikingly, these parallel behavioral responses to a
given CS diverge in their sensitivity to changes in outcome value,
leading to the expression of conditioned responses
—
such as increased
pupil dilation
—
that
fl
exibly adapt to the updated value of an outcome,
and others that persist unchanged, despite the outcome being deva-
lued within the same individual
16
.
Devaluation insensitive behaviors are often suggested to rely on
brain signals approximated through model-free reinforcement learn-
ing algorithms that use reward prediction errors to make predictions
basedoncachedvalues
14
,
46
. A key empirical test of this hypothesis as
applied to Pavlovian conditioning would be whether brain regions
correlating with model-free reinforcement learning based on reward
prediction errors are sensitive to changes in outcome value. If brain
regions sensitive to reward prediction errors are indeed insensitive to
devaluation, this would provide evidence for the role of model-free
reinforcement-learning in the acquisition of devaluation insensitive
Pavlovian behaviors. On the other hand, if such reward prediction
error signal coding brain regions are actually sensitive to outcome
devaluation, this would suggest that Pavlovian reward prediction
error-based learning is not model-free.
Within a model-based framework, some computations would be
expected to be devaluation sensitive while others would not. Model-
based predictive representations of expected-value should be deva-
luation sensitive by de
fi
nition, as these representations are proposed
to emerge by integrating knowledge of stimulus-stimulus associations
with knowledge about current expected outcome value. On the other
hand, internal representations of the cognitive model itself should not
be sensitive to changes in outcome-value, for instance, information
about where in the environment an outcome is expected to occur.
Here, we scanned human participants with fMRI while they per-
formed a Pavlovian learning paradigm
16
, in which they were asked to
learn associations between various neutral images and videos of the
delivery of a food outcome (see Fig.
1
A). There were
fi
ve images: one
image was more often associated with the delivery of sweet food on
the left side of the screen (CS+ left sweet); one image was more often
associated with the delivery of the sweet outcome on the right side of
the screen (CS+ right sweet); one image was more often associated
with the delivery of salty food on the left side of the screen (CS+ left
salty); one image was more often associated with the delivery of the
salty outcome on the right side of the screen (CS+ right salty); and
another image was more often associated with no outcome (CS
−
). We
measured the pupil dilation at the CS onset as a conditioned response
Fig. 1 | Schematic representation of the experimental design. A
Illustration of
the sequence of events in a trial during the acquisition phase administered before
devaluation. At the beginning of each trial a conditioned stimulus (CS) was pre-
sented randomly in the upper or lower portion of the screen for 1.5
–
4.5 s (uniformly
distributed). After an anticipation screen of 3 s, a video showing the snack delivery
appeared either to the right or the left side of the screen for 3 s. Participants were
asked to detect the location of the video of the snack delivery as rapidly as possible.
The intertrial interval (ITI) lasted for 4
–
8 s (uniformly distributed). At the end of
each run, participants received the actual snacks delivered during the task and were
allowed to eat them.
B
Illustration of the sequence of events in a trial during the test
phase administered after devaluation. All aspects were identical to the acquisition
phase with the exception that the outcome delivery happened behind two black
patches.
C
Manipulation check of the outcome devaluation procedure. Mean
pleasantness ratings of the snack that was devalued through the selective satiation
procedure (devalued pleasantness) and the snack that was not (valued pleasant-
ness). Error bars indicate the within--participant s.e.m.
N
=29participants.
Article
https://doi.org/10.1038/s41467-023-43747-5
Nature Communications
| (2023) 14:8057
2
re
fl
ecting affective value
11
,
13
,
16
and the anticipatory gaze direction (left
vs. right) as a conditioned response re
fl
ecting a speci
fi
c perceptual
representation of the outcome (i.e., its spatial location)
16
.Weused
these conditioned responses to
fi
t a model that learns through reward
prediction errors
—
tracking changes in affective value independently of
the perceptual attributes of the outcome
18
—
and a model that learns
through state prediction errors
—
tracking how unexpected a particular
perceptual outcome state is given the previous state independently of
its affective value
5
,
30
.
We identi
fi
ed brain regions involved in learning associations
between the CS and an outcome
’
s affective value as well as other
attributes of an outcome such as its perceptual features, by correlating
the BOLD signal with trial-by-trial reward- and state- prediction errors.
We then tested the sensitivity of these identi
fi
ed regions to outcome
devaluation. State prediction errors were found to carry information
concerning predictions about two perceptual attributes of an out-
come: its taste identity (sweet or salty) and its spatial localization (left
or right). Therefore, to further investigate the representations of pre-
dictions about outcome attributes and their sensitivity to outcome
devaluation, we performed a supplementary analysis. We imple-
mented a multivoxel pattern analysis (MVPA) on the BOLD responses
to the CS onset. We decoded predicted outcome taste identity by
training a classi
fi
er to discriminate between the CS+ sweet and the CS+
salty associated with the outcome delivery to the left side and then
tested its ability to discriminate between the CS+ sweet and the CS+
salty associated with the outcome delivery to the right side. Following
the same logic, we decoded predicted outcome delivery location by
training a classi
fi
er to discriminate between the CS+ left and the CS+
right associated with the sweet outcome and then tested its ability to
discriminate between the CS+ left and the CS+ right associated with
the salty outcome.
Using this approach, we aimed to test for the extent to which brain
regions involved in implementing different learned associations in
Pavlovian conditioning are sensitive to changes in outcome value. We
further aimed to directly test for the applicability of the distinction
between model-based and model-free reinforcement learning as a
means of explaining differences in devaluation sensitivity across these
different Pavlovian associations.
Results
Behavioral results
Pavlovian learning
. During the acquisition phase, we tested whether
pupil dilation and anticipatory gaze direction re
fl
ect patterns of dis-
tinct classes of Pavlovian response as in Zhang et al.
’
sstudy
45
and our
previous study
16
. We expected pupil dilation to follow a value pattern
(all CSs+ different from CS-) and gaze direction to follow a lateralized
pattern (larger dwell time for CSs+ left compared to CSs+ right and the
CS- on the left side of the screen; larger dwell time for CSs+ right
compared to CSs+ left and the CS- on the right side of the screen).
Pupil dilation
. As expected, a planned contrast analysis on the
CS condition (CSs+ left, CSs+ right, CS- with the following weights:
+0.5, +0.5,
–
1) revealed that the pupil was less constricted for CSs+
left and CSs+ right compared to CS- (
β
=
−
0.030,
SE
= 0.011, 95%
CI = [
−
0.053,
−
0.007],
p
=0.016,
BF
10
=3.68;seeFig.
2
A).
Anticipatory gaze direction
.The
fi
rst planned contrast analysis
on the CS condition (CSs+ left, CSs+ right, CS- with the following
A
B
C
50
100
150
200
250
50
100
150
200
250
B
a
s
e
l
i
n
e
-
c
o
r
r
e
c
t
e
d
p
u
p
i
l
d
i
a
m
e
n
t
e
r
(
a
r
b
.
u
n
i
t
s
)
A
n
t
i
c
i
p
a
t
i
o
n
C
S
+
R
A
n
t
i
c
i
p
a
t
i
o
n
C
S
+
L
F
r
e
q
u
e
n
c
y
F
r
e
q
u
e
n
c
y
0
3
***
Fig. 2 | Effects of Pavlovian conditioning and outcome devaluation on eye
behavior. A
Averaged pupil response over time aligned to the conditioned sti-
mulus (CS) onset for the CSs predicting either the delivery of a snack to the left (CS+
L), the delivery of a snack to the right (CS+ R) or no snack delivery (CS-). Shaded
areas indicate the within--participant s.e.m.
B
Heatmaps of the
fi
xation patterns
during the anticipation screen (normalized frequency), after the offset of CS+ L and
of the CS+ R.
C
Devaluation effect calculated as the mean difference of the deva-
luation induced change for the CS valued and the CS devalued (post[valued
—
devalued]
—
pre[valued
—
devalued]) in the pupil response (CS- corrected) and in the
dwell time of the anticipatory gaze direction (CS- corrected). Error bars indicate
95% con
fi
dence interval adapted for within participants design. Statistical sig-
ni
fi
cance was determined by the interaction term (session: pre or post devalua-
tion × CS: value or devalued) in a linear mixed-effects model. Asterisks indicate
statistically signi
fi
cant differences (
β
=0.040,
SE
= 0.008, 95% CI = [0.023, 0.057],
p
< 0.001,
BF
10
=44.77).
N
=29participants.
Article
https://doi.org/10.1038/s41467-023-43747-5
Nature Communications
| (2023) 14:8057
3
weights: +1,-0.5, -0.5), revealed an increased dwell time in the left
region of interest (ROI) after the perception of CSs+ left compared to
CSs+ right and CS-(
β
=
−
0.072,
SE
=0.019, 95%CI=[
−
0.110,
−
0.033],
p
= 0.0010,
BF
10
=85.07; see Fig.
2
B). The second planned contrast
analysis on the CS condition (CSs+ left, CSs+ right, CS- with the fol-
lowing weights: -0.5, +1, -0.5), revealed an increased dwell time in the
right ROI after the perception of CSs+ right compared to CSs+ left and
CS- (
β
=
−
0.079,
SE
=0.020, 95% CI=[
−
0.118,
−
0.039],
p
=0.0005,
BF
10
=53.18;seeFig.
2
B).
Reaction times
. We tested whether participants
’
reaction times to
the outcome delivery were in
fl
uenced by Pavlovian predictions about
(a) the outcome lateralization (i.e., left or right) and (b) the taste
identity of the outcome (i.e., sweet or salty). For outcome lateraliza-
tion, results showed that participants had a signi
fi
cantly longer reac-
tion time when the side of the outcome (but not the identity) was
different than the one most often predicted by the CS compared to
when it was the same (e.g.,
unexpected side effect
;
β
=
−
0.054,
SE
=
0.012, 95% CI = [
−
0.078,
−
0.0306],
p
< 0.001,
BF
10
= 270.42). For out-
come identity, we did not
fi
nd statistically signi
fi
cant effects, although
descriptively participants showed longer reaction times when the
identity of the outcome (but not the side) was different than the one
most often predicted by the CS compared to when it was the
same (e.g.,
unexpected identity effect
;
β
=
−
0.017,
SE
= 0.010, 95% CI =
[
−
0.037, 0.002],
p
=0.101,
BF
10
=0.579).
Outcome devaluation
. A statistically signi
fi
cant interaction between
session (pre- or post-satiation) and outcome (valued or devalued)
showed that the outcome devaluation procedure decreased the plea-
santness of the devalued food outcome in comparison to the valued
food outcome (
β
= 0.629,
SE
= 0.113, 95% CI = [0.406, 0.869],
p
> 0.001,
BF
10
= 434.23; see Fig.
1
C and Supplementary Fig. 4).
Outcome devaluation induced changes
. To test for sensitivity to
outcome value, we compared the change induced by devaluation in
the differential conditioned responses (i.e.,
CS
+
−
CS
−
) to the still
valued CS+ to the change induced by devaluation to the devalued CS+
in a 2 (session: pre- or post-satiation) by 2 (CS: valued or devalued)
interaction. We expected the conditioned pupil response to adapt
more readily to outcome devaluation than the conditioned antici-
patory gaze direction.
Pupil dilation
.WeaveragedthepupilresponseoftheCSsasso-
ciated with the valued outcome and the CSs associated with the
devalued outcome and corrected it by subtracting the average pupil
dilation during the CS-. We did this operation at two time points: the
last run before satiation and the test run. A statistically signi
fi
cant
interaction between session and CS showed that the decrease in pupil
dilation induced by satiation was larger for the CSs associated with the
devalued outcome than the CSs associated with the valued outcome
(
β
=0.040,
SE
= 0.008, 95% CI = [0.023, 0.057],
p
< 0.0001,
BF
10
= 44.77; see Fig.
2
C and Supplementary Fig. 5A).
Anticipatory gaze direction
. We averaged dwell time allocated
to the congruent region of interest (ROI) for all the CSs+ (dwell time in
the right ROI after CSs+ right and dwell time in the left ROI after CSs+
left) for the CSs associated with the devalued outcome (CS devalued)
and the CSs associated with the valued outcome (CS valued) and
corrected it by subtracting the averaged dwell time during the CS- over
both ROI. We did this operation at two time points: the last session
before satiation and the test session. We did not
fi
nd evidence for
an interaction between session and CS (
β
= 0.013,
SE
= 0.007, 95%
CI = [
−
0.0006, 0.027],
p
=0.0710,
BF
10
=0.262; see Fig.
2
CandSup-
plementary Fig. 5B).
Reaction times
. We also measured reaction times taken to guess
which video was being displayed behind the black patches during the
test session following the CS associated with the valued outcome and
the devalued outcome. We did not
fi
nd a statistically signi
fi
cant
difference between the CS valued and the CS devalued conditions
(
β
=0.006,
SE
= 0.008; 95% CI = [
−
0.009, 0.021],
p
= 0.459,
BF
10
= 0.440).
fMRI Results
Parallel Pavlovian predictions about affective value and perceptual
attributes of the outcome
. To identify the brain ROIs separately
involved in implementing Pavlovian predictions about the affective
value and perceptual attributes of the outcome, respectively, we
derived trial-by-trial prediction errors during the
fi
rst two runs from
two models: one learning through reward prediction errors
—
tracking
changes in affective value, independently of the perceptual attributes
of the outcome itself; and the other learning through state prediction
errors
—
tracking how unexpected a particular perceptual outcome
state is independently of its affective value. We then tested for the
sensitivity to devaluation of the ROIs identi
fi
ed with these two models.
Reward prediction errors
. We tested the devaluation sensitivity
of the brain regions involved in rew
ard prediction error coding. To do
so, we de
fi
ned ROIs by extracting the contrast correlating with the trial-
by-trial reward prediction errors
. We focused on three ROIs identi
fi
ed
by this contrast: one ROI covering parts of the ventral striatum and of
the sgACC (VS / sgACC), a second RO
I covering parts of the midbrain,
andathirdROIcoveringpartsofthevmPFC(seeFig.
3
AandTable
1
).
To test for devaluation effects inside these ROIs, we compared
activity while participants expected a valued versus a devalued out-
come, during the run after the devaluation procedure. We also used
pseudo-extinction, whereby the visual presentation of the outcomes
was obscured behind two black patch covers present at the time of
the outcome delivery. Pseudo-extinction is a crucial manipulation
that prevents rapid relearning of a CS
’
s expected value via the
newly devalued outcome. Thus, this procedure allows predictive
representations linked to the incentive value of the predicted
outcome to be dissociated from those associated with outcome-value
insensitive representations. We observed a statistically signi
fi
cant
devaluation effect in the VS / sgACC ROI (
β
=
−
0.149,
SE
=0.057,95%
CI = [
−
0.267,
−
0.030],
p
=0.0157,
BF
10
=2.98;seeFig.
3
B and Supple-
mentary Fig. 6), which survived correction for multiple comparisons
across ROIs. We did not
fi
nd statistical evidence for a devaluation
effect in the midbrain ROI (
β
=
−
0.037,
SE
= 0.054, 95% CI = [
−
0.149,
0.074],
p
= 0.498,
BF
10
= 0.230; see Fig.
3
B) and the vmPFC ROI
(
β
=
−
0.190,
SE
= 0.142, 95% CI = [
−
0.481, 0.099],
p
= 0.189,
BF
10
=
0.390; see Fig.
3
B and Supplementary Fig. 6).
State prediction errors
. We next tested the devaluation sensi-
tivity of the brain regions putatively involved in model-based learning,
such as when forming stimulus
–
stimulus associations between stimuli
and an outcome
’
s perceptual features. To do so, we de
fi
ned ROIs by
extracting the contrast correlating with the trial-by-trial state predic-
tion errors, which tracked how unexpected a particular outcome state
is, given the previous state. We focused on four ROIs identi
fi
ed from
this contrast: one covering parts of the lateral orbitofrontal cortex and
anterior insula (OFC), a second covering parts of the middle frontal
gyrus and inferior frontal gyrus (MFG), a third covering parts of the
superior frontal gyrus (SFG), and a fourth covering parts of the mid-
brain (see Fig.
3
CandTable
2
)
We did not
fi
nd evidence for a statistically signi
fi
cant
effect of devaluation in the MFG ROI (
β
=
−
0.124,
SE
=0.174, 95%
CI = [
−
0.482, 0.233],
p
=0.480,
BF
10
=0.298; see Fig.
3
D and Sup.
Fig. 7), the SFG ROI (
β
=
−
0.031,
SE
= 0.259, 95% CI = [
−
0.455,
0.392],
p
=0.878,
BF
10
=0.258; see Fig.
3
D and Sup. Fig. 7), the
OFC ROI (
β
=
−
0.039,
SE
=0.141, 95% CI=[
−
0.320, 0.241],
p
=0.781,
BF
10
=0.281; see Fig.
3
D and Supplementary Fig. 7), or
the midbrain ROI (
β
=0.0526,
SE
= 0.114, 95% CI = [
−
0.182, 0.287],
p
=0.649,
BF
10
=0.208; see Fig.
3
D and Supplementary Fig. 7)
State prediction errors could potentially be involved in mediating
learning about two different perceptual attributes of an outcome: a
Article
https://doi.org/10.1038/s41467-023-43747-5
Nature Communications
| (2023) 14:8057
4
stimulus could be unexpected because of a violation in the expected
taste identity of the outcome (sweet or salty) or because of an unex-
pected arrival of the outcome in a particular spatial location (left or
right side). To test to what extent these two aspects are re
fl
ected in the
state prediction error brain signals, we extracted the
β
effect of the
state prediction error from the state prediction error ROIs and aver-
aged this across the different ROIs. Then we correlated the averaged
β
effect of the state prediction error against the unexpected side effect
and the unexpected identity effect as measured with reaction times
during the two runs of the acquisition phase. More precisely, to
compute a reaction time index re
fl
ecting the unexpected side effect,
we only used the trials where the identity was the one most often
predicted by the CS but the side of the outcome varied. We subtracted
the average reaction time on trials where the side of the outcome was
thesameastheonemostoftenpredictedbytheCSfromtheaverage
reaction time on trials where the side was different from the one most
often predicted by the CS. To compute a reaction time index re
fl
ecting
the unexpected identity effect, we only used the trials where the side
was the one most often predicted by the CS but the identity of the
outcome varied. We subtracted the average reaction time in trials
where the identity was the same as the one most often predicted by the
CS from the average reaction time on trials where the identity was
different from the one most often predicted by the CS.
We found that the magnitude of state prediction errors was
associated with the magnitude of the unexpected identity effect in
reaction times (
β
=1.956,
SE
= 0.830, 95% CI = [0.339, 3.583],
p
=0.026,
BF
10
= 3.05; see Fig.
4
A) and with the magnitude of the unexpected side
effect in reaction times (
β
= 1.883,
SE
= 0.671, 95% CI = [0.566, 3.199],
p
= 0.009,
BF
10
=7.30; see Fig.
4
A). As a control, we correlated these
behavioral indexes with the
β
effect of reward prediction errors. We
did not
fi
nd conclusive evidence that reward prediction errors are
associated with the magnitude of ei
ther the unexpected taste identity
effect (
β
=0.966,
SE
=0.520, 95% CI=[
−
0.053, 1.986],
p
= 0.066,
BF
10
= 1.40) or the unexpected side effect (
β
=0.587,
SE
=0.441, 95%
CI = [
−
0.277, 1.452],
p
=0.195,
BF
10
= 0.816) in the reaction times
(see Fig.
4
B).
t-score
t-score
Fig. 3 | Reward and state prediction errors and sensitivity to outcome deva-
luation. A
Brain regions correlating with reward prediction error (Reward PE).
Sensitivity to outcome devaluation was esti
mated by calculating the mean difference
between the betas for the valued contrast - betas for devalued contrast in the regions
of interest (ROI).
B
Sensitivity to outcome devaluation in the midbrain ROI the
ventral striatum / sgACC ROI (VS), the vent
romedial prefrontal cortex ROI (vmPFC).
C
Brain regions correlating with state prediction error (State PE).
D
Sensitivity to
outcome devaluation in the midbrain ROI, the superior frontal gyrus ROI (SFG), the
bilateral orbitofrontal/anterior insula ROI
(OFC), the middle prefrontal gyrus/inferior
frontal gyrus ROI (MFG). The valued contrast was de
fi
ned as the difference in the
BOLD signal at the outcome delivery (displayed behind two black patches) after
the perception of the positive conditioned stimulus (CS+) valued versus the negative
conditioned stimulus (CS-). The devalued contrast was de
fi
ned as the difference
in the BOLD signal at the outcome delivery (displayed behind two black patches)
after the perception of the CS+ devalued versus the CS-. Error bars indicate 95%
con
fi
dence interval adapted for within participants design. Statistical signi
fi
cance
was determined by the effect of the outcome value (value or devalued) in a
linear mixed model. Asterisks in
dicate the statistically signi
fi
cant difference that
survives correction for multiple comparisons across ROI (
β
=
−
0.149,
SE
=0.057,95%
CI = [
−
0.267,
−
0.030],
p
=0.0157,
BF
10
=2.98).
N
= 29 participants.
Article
https://doi.org/10.1038/s41467-023-43747-5
Nature Communications
| (2023) 14:8057
5