Clinical Review & Education
Neuroscience and Psychiatry
Mood Instability and Reward Dysregulation—
A Neurocomputational Model of Bipolar Disorder
Liam Mason, PhD, DClinPsy; Eran Eldar, PhD; Robb B. Rutledge, PhD
Classically, bipolar disorder is seen as a relapsing-remitting condi-
tion with episodes of dizzying and capricious highs (manic epi-
sodes) that are clearly separated from melancholic lows and inertia
(depressive episodes) and interspersed with remission. Yet, in the
clinical setting, the picture is often far more complicated. Bipolar
disorder lies on a spectrum with subtypes that are defined by arbi-
trary and often unfulfilled criteria for the severity and duration of
mood episodes, which leads to high rates of “unspecified” diagno-
ses. “Mixed affective” episodes, in which depressive and manic
symptoms co-occur, are the norm rather than the exception. Mood
instability also persists out of episode, causing as much impairment
as within-episode symptoms1 and muddying the classically episodic
presentation.
These complexities challenge existing models of bipolar disorder2
and broader psychobiological frameworks that describe the relation-
ship between mood and behavior. A prominent theory proposes that
bipolar disorder arises from the dysregulation of a ‘“behavioral acti-
vation system” that drives positive affect, confidence, and the ap-
proach of rewarding stimuli,2 and temporary increases and de-
creases in this system result in mania and depression, respectively.
However, a plausible neurobiological mechanism for these tempo-
rary changes is lacking. There is fundamental disagreement about the
core pathophysiology of bipolar disorder. Some research supports
a “reward hypersensitivity” account,3 but other research suggests that
reward sensitivity is fundamentally blunted.4 In contrast to existing
models that do not make quantitative predictions, we argue that a
new computational approach, based on a dynamic understanding of
the 2-way relationship between mood and reward sensitivity, can pro-
vide insight into the mechanisms by which patients with bipolar dis-
order transition between relapse and recovery, as well as explain
mixed affective states and pervasive mood instability.
The application of computational approaches to the study of
emotion and decision making provides new ways to relate momen-
tary changes in mood and behavior to well-defined neural circuits.
Recent studies have found that mood fluctuations are highly con-
tingent on reward prediction error signals that are represented by
activity in the ventral striatum, a region that is involved in goal-
directed behavior.5 These signals encode how rewarding an out-
come was compared with what was expected. For example, posi-
tive surprises (eg, winning £100 [$132] on a scratch card) elicit striatal
activity and a positive mood state. This then biases how subse-
quent rewards are perceived, increasing their perceived value when
mood is elevated (and conversely decreasing perceived value when
mood is low). Expectations that guide future decisions are up-
dated based on these mood-biased reward prediction errors.6,7 A
moderate mood bias helps individuals adapt quickly to an environ-
ment that is changing, either for better or for worse. However, if
mood strongly biases reward perception, as we propose occurs in
jamapsychiatry.com
© 2017 American Medical Association. All rights reserved.
Downloaded From: by a UNIVERSITY OF SYDNEY LIBRARY User on 11/03/2017
bipolar disorder, the result is recursive cycles that cause mood, ex-
pectations, and behavior to escalate to extremes (Figure, A).6 Im-
portantly, the same elevated mood bias parameter leads to hyper-
sensitivity to rewards when mood is high and hyposensitivity to
rewards when mood is low, reconciling previously contradictory ac-
counts of bipolar disorder.3,4 Empirical support for the model comes
from the finding that trait differences in propensity toward el-
evated moods were explained by the degree to which mood induc-
tion biased reward valuation, as inferred from changes in striatal re-
sponses and preference for options encountered when mood was
elevated (Figure, B).6
Our neurocomputational model makes several predictions about
how mood changes will escalate, resolve, and reverse in bipolar dis-
order. We predict that a stronger mood bias parameter that ampli-
fies normal mood fluctuations in response to rewards will be asso-
ciated with less time in remission and more severe mood episodes
(Figure, C). We also predict that high mood bias parameters are a
risk factor for developing bipolar disorder, consistent with findings
that this parameter correlates with trait measures of bipolar disor-
der vulnerability (Figure, B).6 The mood bias parameter, being stable
for a given individual, should also be similar across euthymic, de-
pressed, and manic phases. Through amplification of normal mood
fluctuations, the model also accounts for the mood instability that
occurs both during episodes and residually between mood epi-
sodes. For example, as mood episodes progress, expectations be-
come increasingly discrepant with the objective value of out-
comes. Surprises that run counter to the valence of the episode
become increasingly likely, causing mood and behavior to fluctu-
ate (Figure, C). This greater variability in prediction errors might also
account for mood lability in mixed affective states, which could be
readily tested using computational modeling.
Another parameter within the model is the learning rate, which
determines how quickly expectations are updated based on re-
ward prediction errors. Learning rate asymmetries may explain the
dominance of a positive or negative mood (eg, in patients with uni-
polar mania and bipolar II) (Figure, C). Further, an increasing learn-
ing rate may explain why episode frequency can increase across the
course of the disorder, at least for some patients (“kindling”). In a
changeable environment, it is advantageous to increase the rate of
learning to adapt behavior.5 The volatility that results from re-
peated relapses and their sequelae could strengthen the belief that
further changes are likely, warranting an increased learning rate to
adapt to these changes. However, an increased learning rate would
also increase spontaneous mood instability, increasing the likeli-
hood of further mood episodes. Alternatively, a gradual increase in
mood bias parameters could also lead to kindling. Testing which pa-
rameter provides a better explanation (eg, by comparing model para-
meters among patients with bipolar disorder in early and later stages)
(Reprinted) JAMA Psychiatry Published online October 11, 2017 E1
 Clinical Review & Education Neuroscience and Psychiatry
Figure. Neurocomputational Model of Mood Instability
A Positive feedback cycles underlying mood instability
Expectations
become Manic Cycle
overly low
Biased
perception Higher
mood
Positive
surprises
Negative
Lower surprises
mood Biased
perception
Expectations
become overly high
Depressive Cycle
C Computer simulations for different mood bias parameters
Normal mood
variation Cyclothymia Bipolar IIa
Increased mood bias
is important because these explanations imply different underly-
ing neural circuitry and interventions.
Model parameters provide a way of quantifying mood distur-
bances in a mechanistic way. By quantifying mood and behavior in
relation to well-defined neural circuits, this computational ap-
proach strongly accords with the dimensional approach that is ad-
vocated by the Research Domain Criteria framework. Although the
ARTICLE INFORMATION
Author Affiliations: Max Planck University College
London Centre for Computational Psychiatry and
Ageing Research, University College London, London,
England (Mason, Eldar, Rutledge); Wellcome Trust
Centre for Neuroimaging, University College
London, London, England (Eldar, Rutledge).
Corresponding Author: Liam Mason, PhD,
DClinPsy, Max Planck University College London
Centre for Computational Psychiatry and Ageing
Research, University College London, London,
England (l.mason@ucl.ac.uk).
Published Online: October 11, 2017.
doi:10.1001/jamapsychiatry.2017.3163
Conflict of Interest Disclosures: None reported.
Funding/Support: Dr Rutledge receives research
support from a Career Development Award from
the Medical Research Council. Dr Eldar receives
research support from grant 095844/Z/11/Z from
the Wellcome Trust Cambridge–University College
London Mental Health and Neurosciences Network.
The Max Planck University College London Centre is
a joint initiative supported by University College
E2 JAMA Psychiatry Published online October 11, 2017 (Reprinted)
© 2017 American Medical Association. All rights reserved.
Downloaded From: by a UNIVERSITY OF SYDNEY LIBRARY User on 11/03/2017
B Effects of mood on reward valuation
1.0
Increase in Reward Response
0.5
0 as better than they are, increasing
−0.5
−1.0
−1.5
1.5
Bipolar I
London and the Max Planck Society. No other
disclosures were reported.
Role of the Funder/Sponsor: The funding
organizations had no role in the preparation,
review, or approval of the manuscript or decision to
submit the manuscript for publication.
Additional Contributions: We thank Peter Dayan,
PhD, Gatsby Computational Neuroscience Unit,
University College London, for helpful comments
and discussions arising from an earlier version of
this paper. He was not compensated for his
contribution.
REFERENCES
1. Broome MR, Saunders KEA, Harrison PJ,
Marwaha S. Mood instability: significance,
definition and measurement. Br J Psychiatry. 2015;
207(4):283-285. doi:10.1192/bjp.bp.114.158543
2. Urosević S, Abramson LY, Harmon-Jones E, Alloy
LB. Dysregulation of the behavioral approach
system (BAS) in bipolar spectrum disorders: review
of theory and evidence. Clin Psychol Rev. 2008;28
(7):1188-1205. doi:10.1016/j.cpr.2008.04.004
Positive mood induction
Negative mood induction
9
0.6 A, Increases in mood follow from
r=
unexpected rewards, which cause
subsequent rewards to be perceived
future expectations and further
escalating mood (manic cycle). Over
r= time, expectations become
−0.
45 increasingly discrepant with actual
outcomes, with the resulting
negative surprises setting in motion a
2.0 2.5 3.0 3.5 4.0 depressive cycle. B, Vulnerability to
Trait Mood Instability bipolar disorder is associated with a
greater influence of the current mood
on striatal response to rewards.6
Expectation C, Stronger mood bias parameters
Prediction error inflate reward prediction errors,
Mania Mood causing expectations and mood to
hypomania oscillate, akin to mood episodes
(shaded area).
Depression
a
A negative mood predominates if
the learning rate is higher for
negative than positive reward
predictions errors.
proposed model accounts for how positive and negative life events
increase the likelihood of mood episodes, the etiology of individual
differences in this parameter remains unaddressed. In addition to
illuminating multiple poorly understood features of bipolar disor-
der, this computational approach provides an inroad for improving
the diagnosis and treatment of mood instability across psychiatric
disorders.
3. Mason L, O’Sullivan N, Montaldi D, Bentall RP,
El-Deredy W. Decision-making and trait impulsivity
in bipolar disorder are associated with reduced
prefrontal regulation of striatal reward valuation.
Brain. 2014;137(pt 8):2346-2355.
4. Schreiter S, Spengler S, Willert A, et al. Neural
alterations of fronto-striatal circuitry during reward
anticipation in euthymic bipolar disorder. Psychol
Med. 2016;46(15):3187-3198. doi:10.1017
/S0033291716001963
5. Eldar E, Rutledge RB, Dolan RJ, Niv Y. Mood as
representation of momentum. Trends Cogn Sci.
2016;20(1):15-24. doi:10.1016/j.tics.2015.07.010
6. Eldar E, Niv Y. Interaction between emotional
state and learning underlies mood instability. Nat
Commun. 2015;6:6149. doi:10.1038/ncomms7149
7. Rutledge RB, Moutoussis M, Smittenaar P, et al.
Association of neural and emotional impacts of
reward prediction errors with major depression.
JAMA Psychiatry. 2017;74(8):790-797.
jamapsychiatry.com