University of ZurichZurich Open Repository and Archive

Winterthurerstr. 190

CH-8057 Zurich

http://www.zora.uzh.ch

Year: 2009

A common role of insula in feelings, empathy and uncertainty

Singer, T; Critchley, H D; Preuschoff, K

Singer, T; Critchley, H D; Preuschoff, K (2009). A common role of insula in feelings, empathy and uncertainty.Trends in Cognitive Sciences, 13(8):334-440.Postprint available at:http://www.zora.uzh.ch

Posted at the Zurich Open Repository and Archive, University of Zurich.http://www.zora.uzh.ch

Originally published at:Trends in Cognitive Sciences 2009, 13(8):334-440.

Singer, T; Critchley, H D; Preuschoff, K (2009). A common role of insula in feelings, empathy and uncertainty.Trends in Cognitive Sciences, 13(8):334-440.Postprint available at:http://www.zora.uzh.ch

Posted at the Zurich Open Repository and Archive, University of Zurich.http://www.zora.uzh.ch

Originally published at:Trends in Cognitive Sciences 2009, 13(8):334-440.

1

24/01/09

A common role of insula in feelings, empathy and uncertainty

Tania Singer1,2, Hugo D. Critchley3, Kerstin Preuschoff2

1Laboratory for Social and Neural Systems Research, University of Zurich, Switzerland

2Institute for Empirical Research in Economics, University of Zurich, Switzerland

3Clinical Imaging Sciences Centre, Brighton and Sussex Medical School, Universities of

Brighton and Sussex, UK

Type of Article: Opinion Paper

Corresponding Author:

Prof. Dr. Tania Singer Laboratory for Social and Neural Systems Research Institute for Empirical Research in Economics University of Zurich Bluemlisalpstrasse 10 CH-8006 Zurich E-Mail: singer@iew.uzh.ch

2

Abstract

Although accumulating evidence highlights a crucial role of the insular cortex in feelings,

empathy and processing uncertainty in the context of decision making, neuroscientific models

of affective learning and decision making have mostly focused on structures such as the

amygdala and the striatum. Here we propose a unifying model in which insula cortex supports

different levels of representation of current and predictive states allowing for error-based

learning of both feeling states and uncertainty. This information is then integrated in a general

subjective feeling state which is modulated by individual preferences such as risk aversion

and contextual appraisal. Such mechanisms could facilitate affective learning and regulation

of body homeostasis, and could also guide decision making in complex and uncertain

environments.

Words: 119

3

Introduction

Human insular cortex is a large and highly interconnected structure embedded deep in the

brain (see Box 1). Despite neuroimaging observations of insula engagement across multiple

tasks, the contribution of this region has been largely neglected within neuroscientific theories

and models. The emergence of affective and social neuroscience has accumulated evidence

implicating insular cortex, particularly its most anterior portion (the anterior insula, AI), in

visceral representation and emotional experience. Commentaries emphasize the activation of

AI by motivationally important changes in bodily states as diverse as autonomic arousal,

sensual touch, air hunger, taste, craving and pain1-6. Furthermore, AI activity reflects the

subjective intensity of both one’s own and others’ emotional experiences, for example, when

one experiences distress, pain or disgust7 or when one empathizes with others who are in

these emotional states8-10. Finally, neuroeconomics which studies motivational decision

making, has recently strongly linked AI activity to processing, representing and learning

information about risk and uncertainty11-14.

Until now, however, no model of insula/AI function has integrated its contribution to

physiological and emotional states with empathic understanding or behavioural risk

processing. Here, we propose a novel account of the role of insula/AI in which predictions

and realizations of bodily and affective states are integrated with predictions and realizations

of uncertainty, to engender an integrated feeling state that is shaped by individual risk

preference and appraisal of the context. Before outlining our model and its implications, we

will first summarise relevant studies illustrating insula/AI involvement in: a) bodily

awareness, b) self-related and empathic feelings, c) risk and uncertainty processing (for earlier

extensive reviews of the literature in these respective domains, see for example Refs. 1,2,4,15-

19).

[ INSERT BOX 1 ]

4

The interoceptive cortex: Insula and feelings

Insular cortex is broadly acknowledged as viscerosensory cortex, and implicated in mapping

internal bodily state (including pain and taste) and in representing emotional arousal and

feelings1,2,4. Early ‘peripheral’ theories of emotion argue that visceral and somatic changes

within the body underlie emotional feeling states and provide emotional colour to

perception20: Thus, when exposed to a spider, a perceived change in one’s heart beat can

trigger the experience of fear. Damasio21 and colleagues revived neurobiological interest in

these theories suggesting that insular cortex plays a role in mapping visceral states that are

associated with emotional experience, giving rise to conscious feelings (see Box 2). This

notion was recently extended by Craig who argued that sensory information about the body’s

physiological state is mapped in insula and re-represented in AI (particularly in right AI)

where it becomes consciously accessible, enabling a subjective affective experience or feeling

state1,2. Accordingly, enhanced AI activity is observed during high physiological arousal and

during declarative awareness of changes in bodily states4. Differences in AI size and

reactivity is linked to awareness of bodily responses (such as one’s heartbeat) and to the

general experience of anxiety symptoms22, suggesting that the link between emotional

feelings and conscious perception of bodily response20,21,23 is strongest for anxiety (see Figure

1, Panels A and B).

[ INSERT BOX 2]

In a neurobiological model of anxiety, Paulus and Stein24 suggested that insula computes an

‘interoceptive prediction error’, signalling mismatch between actual and anticipated bodily

arousal, evoking subjective anxiety and avoidance behaviour. This interoceptive signal is

integrated with other information processing streams, such that AI activation is engaged by

5

the joint processing of internal arousal and conscious representation of threat (relevant to ‘two

stage’ models of emotion23, Box 2) and enhanced in anxiety-prone individuals during

emotion- or risk-processing tasks13,25. Moreover, when ‘interoceptive mismatch’ is induced

experimentally using false physiological feedback, enhanced AI activity correlates with

increased emotional salience attributed to previously unthreatening neutral stimuli26.

These findings extend the role of insula/AI cortex beyond representation of current

physiological or emotional feeling states and suggest a role in behaviourally-relevant

computation of predictive states and prediction errors. Also, they suggest through anxiety a

connection between feelings and risk, a crucial point for the model described below.

[ INSERT FIGURE 1]

The interoceptive cortex and its role in empathy

Independent evidence linking insular cortex to predictive feeling states also stems from

research in social neuroscience, focusing on the neural basis of empathy – our ability to share

and understand the emotional states of others (reviewed in Refs. 16-19). Functional

neuroimaging studies demonstrate overlapping patterns of brain activity during the subjective

and the vicarious experience of emotions or sensations. For example, bilateral AI activation is

consistently observed both when one experiences pain and when one observes pain in other

people (see Figure 1, Panel B). Such empathic engagement of AI also reflects dispositional

empathy scores and individual trial-by-trial empathy ratings8-10,27,28 (reviewed in Ref. 16).

Bilateral AI is similarly engaged when one tastes or sees other people taste, unpleasant and,

notably, pleasant drinks8.

6

Singer and others10 proposed a ‘dual function’ for insular representations of bodily states:

First, the primary mapping of internal states forms the basis for predictions of physiological

reactions to emotional stimuli with respect to the self (subjective feeling states). Second, these

predictive representations permit the simulation of how the same emotional stimuli feel to

others (empathic feeling states).

This account predicts that impaired access to one’s own emotional state will cause empathic

deficits. Alexithymia describes the difficulty in identifying and describing emotional feelings,

despite a basic awareness of bodily sensations and arousal. When alexithymic individuals

focus on their emotions, the failure to engage AI reflects both the degree of alexithymia and

deficiencies in trait empathy29. Further studies of empathic brain responses in alexithymic

patients may confirm that deficits in empathy emerge from a failure to simulate forward

representations of bodily states within AI.

This research implicates insula/AI in the representation of predictive feeling states used both

to anticipate emotional events impacting one’s own body and to understand the emotional

experience of others. It also points to a central role of insular cortex in vicarious learning.

The role of anterior insula in uncertainty and uncertainty prediction

Research associated to Neuroeconomics, which explores the neural basis of motivational

decision making, suggest that AI processes and learns about risk and uncertainty (see Box 3

for definition of risk and uncertainty; for simplicity, we use the terms interchangeably).

‘Uncertainty’ describes the inability to fully predict an outcome. Most organisms are sensitive

to uncertainty. For instance, someone may forego possible large rewards in favour of smaller,

but less uncertain, rewards (risk aversion). Imaging studies report increased AI signals when

making risky decisions (gambles) compared to safe decisions13 and in response to increasing

7

task ‘instability’30, complexity31 and ambiguity11, which engender uncertainty through lack of

knowledge. Mathematical tools derived from economics permit a mechanistic description of

how AI represents risk and uncertainty: In a decision-making task, if the probability and

magnitude of each potential outcome are known, the precise risk of each decision can be

calculated. This methodology does not require a subjective evaluation of risk, yet explains

risk-taking behaviour across species32.

INSERT BOX 3: What is risk and how is it assessed?

Preuschoff and co-workers33 applied this model to examine neural substrates of risk

processing: Using a gambling task, activity within bilateral AI was observed to encode the

computed risk prediction when waiting for the outcome of a risky decision and to reflect the

risk prediction error once the outcome was known. This error signal compares predicted risk

to realized risk34 (see Figures 1, Panel D). Through these two levels of representations of

uncertainty information, AI can guide choice selection in risk-sensitive individuals and

modulate learning rates in uncertain environments. Behaviourally, changes in learning rates

correlate with changing risk prediction35. Moreover the level of activity within AI before a

choice is made can predict both risk-taking behaviour and errors resulting from risk-taking12.

These studies suggest that AI does not only process and learn about feeling states but also

process and learn about risk associated with current decisions and is likely to mediate

behavioural and physiological effects of risk prediction.

Towards an integrative model

A comprehensive explanatory account of AI function must be able to link these three lines of

evidence. We propose a model in which insular cortex integrates external sensory and internal

8

physiological signals with computations about their uncertainty. Through AI, this integration

is expressed as a dominant feeling state that modulates social and motivational behaviour in

conjunction with bodily homeostasis. This model is schematically depicted in Figure 2.

[INSERT FIGURE 2]

The model differentiates several mechanisms and levels of representations within insular

cortex (Figure 2A). First, there are modality-specific feeling states: A current feeling state

(how something feels now), a predictive feeling state (how something will feel) and a feeling

prediction error (how wrong the prediction was). In the case of pain, predictive feeling states

simulate the feeling in response to nociceptive stimulation before it is applied to me or others.

This predictive feeling is compared to the actual (current) feeling during stimulation.

Learning from this prediction error (e.g., through reinforcement learning) will result in more

accurate predictive feeling states and fewer prediction errors36-38.

Second, we learn about uncertainty via a parallel mechanism that involves a corresponding set

of representations: actual uncertainty, predictive uncertainty and uncertainty prediction errors

(greater or less uncertainty than predicted) (see Box 3 and Refs. 33,38). In the case of pain, this

mechanism provides a measure of how uncertain I am about the upcoming pain stimulus

within the current environment. It reflects the variance of the pain stimulus (rather than its

mean), in other words, how certain I am about the occurrence and magnitude of the painful

stimulus and the resulting feeling state and – after pain delivery – how (in)accurate my

certainty about the feeling state was.

Third insular cortex integrates information from modality-specific feeling states and

uncertainty with individual preferences (e.g., risk aversion or sensation seeking) and

9

contextual information (e.g., personal belief about the source of the pain) to produce a global

feeling state (see Figure 2B). As decisions between equally expected rewards are biased by

the degree of uncertainty, uncertainty itself possesses motivational value32 and can be

reflected in a feeling state. Bodily and affective responses to uncertain stimuli may facilitate

(and therefore accelerate) behavioural responses aimed at avoiding uncertainty. Accordingly,

it is adaptive for a body to prepare a flight response in highly uncertain situations.

Finally, we suggest that the integration of uncertainty with bodily, affective and sensory

information within AI improves learning and guides decision making alongside with

physiological reactivity. The anatomical interconnections of insular cortex with subcortical

and cortical areas support such higher-level integrated feeling states (see Box 1 and Ref. 2).

Current evidence and model prediction

Our model proposes the existence of predictive feeling states, current feeling states and

feeling state prediction errors that support the error-based learning of feeling states in insula

(Figure 2A). Along with the analogous uncertainty representations, these are likely in

spatially distinct but close neural circuits in insula.

Studies of pain and pleasant touch suggest that, across insular cortex, primary representations

are topographically distinct from predictive and empathic representations: Actual subjective

experience of pain or touch engages mid and posterior insula, while anticipation or

empathizing with others’ experience of pain or touch engages AI1,2,6,10,39. While this is

consistent with a shared representation within AI underlying predictive and empathic feeling

states in the domain of pain (see above and Ref. 10), the caudo-rostral organization of

representational levels within the insula may not generalize to other sensory modalities,

including taste, where primary cortex lies within AI8. Evidence for prediction errors within AI

10

have been reported both for learning about nociceptive stimuli40 and, recently, for learning

about uncertainty33,34.

Our model further proposes that a dominant subjective feeling state arises in insula from

integrating modality-specific feeling states about self and others (e.g., pain, touch) with

information about uncertainty as well as individual preferences (Figure 2B).

Although research has not focused on the simultaneous manipulation of uncertainty, bodily

signals and feeling states, several studies suggest direct interaction between the three

domains. For example, behaviourally, physically identical pain stimuli are rated as being more

painful when they are unpredictable rather than predictable. Interestingly, the unpredictability

of pain is reflected in anterior insula41,42 and the predictability of pain is reflected in posterior

insula activity41. Misleading anticipatory cues also modulate perceived pain through

engagement of mid- and posterior insula, in other words, influence the representation of the

subjective experience of the stimulus43. Furthermore, correlations have been reported between

risk prediction and the physiological arousal of financial traders, between risk perception and

dispositional fear as well as between risk-taking behaviour, anxiety32 and harm avoidance13.

In addition, patients with insular damage show a reduced sensitivity to risk44 and impaired

emotional intelligence45. Interestingly, Simmons et al.46 demonstrated an interaction between

emotion and uncertainty in mid insula: The neural signal increased with increasing intolerance

of uncertainty during an affective ambiguity task, but not during a non-affective ambiguity

task.

Finally, we suggest that the integration of uncertainty with bodily, affective and sensory

information within AI influences decision making in conjunction with homeostatic body

regulation. Several recent studies support such a link. During decision making, anticipatory

11

right AI activation predicts low risk choices, while overall insula activation predicts risk

preferences12,47. Sanfey and colleagues48 revealed that anterior insula activation precedes the

choice to defect against unfair but not fair players and Naqvi et al.5 found that smokers with

brain damage involving the insula, compared to smokers with lesions in other areas, were

more likely to quit smoking easily and immediately, without relapse, and without persistence

of the urge to smoke, a finding that also points to the important role of insula in craving and

addiction. In addition, patients with insular damage demonstrate poor decision making during

gambling experiments resulting in higher losses44. Finally, it is interesting to note that during

instrumental conditioning, rats with lesions to the insular cortex fail to adjust their choice

behaviour to variations in motivational state during learning49.

The above studies strongly support a role for insular cortex in learning predictive feeling

states and mediating the interaction between feeling states and uncertainty. However, no

studies have explored the interaction between different feeling states (predictive, current and

error) and uncertainty during learning. A clever combination of previous paradigms can

provide further insight. Consider a conditioning paradigm using pain40 in which the intensity

(low vs. high) and the probability of stimulus delivery (50% vs. 100%) are varied

simultaneously. Participants in two pre-selected groups (risk-averse vs. risk-seeking) rate how

they feel after each trial. Our model predicts spatially distinct signals in AI reflecting

predictions and prediction errors related to the intensity of pain on the one hand and

uncertainty on the other. The model predicts different activation patterns for different

modalities (e.g., pain vs. taste). However, uncertainty signals in AI could be shared across

modalities. Another peak of activation in AI should reveal the integrated feeling state

reflecting the interaction between risk preferences, pain and uncertainty: Risk-averse

participants will feel more negative about the painful stimulus if it is accompanied by the

negative feeling of situational uncertainty (50% predictability). In contrast, for a risk-seeking

12

participant, the negative feeling about the painful stimulus will be alleviated by the positive

feeling towards uncertainty.

There is less evidence about the effects of uncertain environments on empathy. Recent studies

highlighted the sensitivity of empathic responses within AI to contextual factors including

affective bond or the perceived source for others’ pain (for an overview see19). The

trustworthiness of others47,50 could be used experimentally to influence the degree of empathic

uncertainty. Similarly, vicarious aspects of emotional learning and feeling prediction errors

could be explored by providing feedback to empathizing participants about the subjective

feeling or physiological states of others receiving pain. If this information does not match the

empathic prediction (what the participant thought the other was feeling), a vicarious

prediction error signal should be observed within AI.

Conclusion

Affective neuroscience, social neuroscience and neuroeconomics demonstrate that insular

cortex, particularly AI, is involved in processing subjective feelings, empathy and uncertainty.

We propose an account of insular cortex function in which sensory, affective and bodily

information is integrated with information about uncertainty to generate a dominant subjective

feeling state. Learning about feeling states is supported by insular representations of current

feeling states, predicted feeling states and feeling prediction errors. Learning about

uncertainty is supported by analogous representations in anterior insula. A dominant

integrated feeling state includes current and predictive feeling states and uncertainty as well as

individual preferences. It is putatively experienced as emotional confidence that serves to

motivate and guide social interaction and decision-making behaviour.

13

Extending the computational model of insula during risk-taking to include emotion and

empathy could considerably advance our understanding of affective processes and emotion

learning, with particular relevance to understanding decision making in complex social

environments. Conversely, theories of economic decision making may benefit from

understanding the role of bodily signals, feeling states and subjectivity in shaping complex

decisions.

INSERT QUESTION BOX: Open Questions

Acknowledgements

We thank Klaas Enno Stephan and Peter Bossaerts for useful comments on the present paper. HDC was supported by a programme grant from the Wellcome Trust.

14

Reference List 1 Craig,A.D. (2003) Interoception: the sense of the physiological condition of the body. Current Opinion in Neurobiology 13, 500-505

2 Craig,A.D. (2009) How do you feel -- now? The anterior insula and human awareness. Nat Rev Neurosci 10, 59-70

3 Critchley,H.D. et al. (2002) Fear conditioning in humans: the influence of awareness and autonomic arousal on functional neuroanatomy. Neuron 33, 653-663

4 Critchley,H.D. (2005) Neural mechanisms of autonomic, affective, and cognitive integration. J. Comp. Neurol. 493, 154-166

5 Naqvi,N.H. et al. (2007) Damage to the insula disrupts addiction to cigarette smoking. Science 315, 531-534

6 Ploghaus,A. et al. (1999) Dissociating pain from its anticipation in the human brain. Science 284, 1979-1981

7 Wicker,B. et al. (2003) Both of us disgusted in My insula: the common neural basis of seeing and feeling disgust. Neuron 40, 655-664

8 Jabbi,M. et al. (2007) Empathy for positive and negative emotions in the gustatory cortex. Neuroimage 34, 1744-1753

9 Saarela,M.V. et al. (2007) The compassionate brain: humans detect intensity of pain from another's face. Cereb Cortex 17, 230-237

10 Singer,T. et al. (2004) Empathy for pain involves the affective but not sensory components of pain. Science 303, 1157-1162

11 Huettel,S. et al. (2006) Neural signatures of economic preferences for risk and ambiguity. Neuron 49, 765-775

12 Kuhnen,C.M. and Knutson,B. (2005) The neural basis of financial risk taking. Neuron 47, 763-770

13 Paulus,M.P. et al. (2003) Increased activation in the right insula during risk-taking decision making is related to harm avoidance and neuroticism. Neuroimage 19, 1439-1448

14 Preuschoff,K. et al. (2008) Human insula activation reflects risk prediction errors as well as risk. J. Neurosci. 28, 2745-2752

15 Platt,M.L. and Huettel,S.A. (2008) Risky business: the neuroeconomics of decision making under uncertainty. Nat Neurosci 11, 398-403

16 Singer,T. and Leiberg,S. (2009) Sharing the emotions of others: The neural bases of empathy. In The Cognitive Neurosciences IV (Gazzaniga,M.S., ed), pp. 971-984, MIT Press

17 Decety,J. and Lamm,C. (2006) Human empathy through the lens of social neuroscience. TheScientificWorldJOURNAL 6, 1146-1163

15

18 Keysers,C. and Gazzola,V. (2006) Towards a unifying neural theory of social cognition. Progress in Brain Research 156, 379-401

19 Singer,T. and Lamm,C. (2009) The social neuroscience of empathy. The Year in Cognitive Neuroscience 2009: Annals of the New York Academy of Sciences 1156, 81-96

20 James,W. (1894) The physical basis of emotion. Psychological Review 1, 516-529

21 Damasio,A. (1994) Descartes' error and the future of human life. Scientific American 271, 144

22 Critchley,H.D. et al. (2004) Neural systems supporting interoceptive awareness. Nature Neuroscience 7, 189-195

23 Schachter,S. and Singer,J.E. (1962) Cognitive, social, and physiological determinants of emotional state. Psychological Review, 379-399

24 Paulus,M.P. and Stein,M.B. (2006) An insular view of anxiety. Biological Psychiatry 60, 383-387

25 Stein,M.B. et al. (2007) Increased amygdala and insula activation during emotion processing in anxiety-prone subjects. Am J Psychiatry 164, 318-327

26 Gray,M.A. et al. (2007) Modulation of emotional appraisal by false physiological feedback during fMRI. PLoS ONE 2, e546

27 Singer,T. et al. (2006) Empathic neural responses are modulated by the perceived fairness of others. Nature 439, 466-469

28 Singer,T. et al. (2008) Effects of oxytocin and prosocial behavior on brain responses to direct and vicariously experienced pain. Emotion 8, 781-791

29 Silani,G. et al. (2008) Levels of emotional awareness and autism: an fMRI study. Social neuroscience 3, 97-112

30 Elliott,R. et al. (2000) Dissociable neural responses in human reward systems. J Neurosci 20, 6159-6165

31 Grinband,J. et al. (2006) A neural representation of categorization uncertainty in the human brain. Neuron 49, 757-763

32 Weber,E.U. et al. (2004) Predicting risk sensitivity in humans and lower animals: risk as variance or coefficient of variation. Psychological Review 111, 430-445

33 Preuschoff,K. et al. (2008) Human insula activation reflects risk prediction errors as well as risk. J Neurosci 28, 2745-2752

34 De Martino,B. et al. (2009) The neurobiology of reference-dependent value computation. J. Neurosci. 29, 3833-3842

35 Behrens,T.E.J. et al. (2007) Learning the value of information in an uncertain world. Nature Neuroscience 10, 1214-1221

16

36 Sutton,R.S. and Barto,A.G. (1981) Toward a modern theory of adaptive networks: expectation and prediction. Psychological Review 88, 135-170

37 Schultz,W. et al. (1997) A neural substrate of prediction and reward. Science 275, 1593-1599

38 Preuschoff,K. and Bossaerts,P. (2007) Adding prediction risk to the theory of reward learning. Ann N Y Acad Sci 1104, 135-146

39 Lovero,K.L. et al. (2009) Anterior insular cortex anticipates impending stimulus significance. Neuroimage 45, 976-983

40 Seymour,B. et al. (2004) Temporal difference models describe higher-order learning in humans. Nature 429, 664-667

41 Carlsson,K. et al. (2006) Predictability modulates the affective and sensory-discriminative neural processing of pain. Neuroimage 32, 1804-1814

42 Brown,C.A. et al. (2008) Modulation of pain ratings by expectation and uncertainty: Behavioral characteristics and anticipatory neural correlates. Pain 135, 240-250

43 Nitschke,J.B. et al. (2006) Altering expectancy dampens neural response to aversive taste in primary taste cortex. Nat Neurosci 9, 435-442

44 Clark,L. et al. (2008) Differential effects of insular and ventromedial prefrontal cortex lesions on risky decision-making. Brain 131, 1311-1322

45 Bar-On,R. et al. (2003) Exploring the neurological substrate of emotional and social intelligence. Brain 126, 1790-1800

46 Simmons,A. et al. (2008) Intolerance of uncertainty correlates with insula activation during affective ambiguity. Neuroscience Letters 430, 92-97

47 Engelmann,J.B. et al. (2009) Expert financial advice neurobiologically "Offloads" financial decision-making under risk. PLoS ONE 4, e4957

48 Sanfey,A.G. et al. (2003) The neural basis of economic decision-making in the ultimatum game. Science 300, 1755-1758

49 Balleine,B.W. and Dickinson,A. (1998) Goal-directed instrumental action: contingency and incentive learning and their cortical substrates. Neuropharmacology 37, 407-419

50 Behrens,T.E.J. et al. (2008) Associative learning of social value. Nature 456, 245-249

51 Mesulam,M.M. and Mufson,E.J. (1982) Insula of the old world monkey. III: Efferent cortical output and comments on function. J. Comp. Neurol. 212, 38-52

52 Mesulam,M.M. and Mufson,E.J. (1982) Insula of the old world monkey: Architectonics in the insulo-orbito-temporal component of the paralimbic brain. J. Comp. Neurol. 212, 1-22

53 Flynn,F.G. et al. (1999) Anatomy of the insula - functional and clinical correlates. Aphasiology 13, 55-78

17

54 Ongür,D. et al. (2003) Architectonic subdivision of the human orbital and medial prefrontal cortex. J. Comp. Neurol. 460, 425-449

55 Scherer,K.R. (1984) On the nature and function of emotion: A component process approach, Erlbaum

56 Critchley,H.D. et al. (2000) Neural activity relating to generation and representation of galvanic skin conductance responses: a functional magnetic resonance imaging study. The Journal of Neuroscience 20, 3033-3040

18

Box 1: Anatomy and connectivity of insular cortex

Insular cortex is located bilaterally within the brain, tucked away under the posterior part of

the frontal lobe and the anterior part of the temporal lobe. It connects anteriorly to the lateral

orbitofrontal cortex (OFC) and posteriorly to the superior temporal cortex. Its main division is

the central insular sulcus that separates anterior from posterior insula although more recent

work also focuses on the dissociation of ventral and dorsal parts.

The dense anatomical connectivity of anterior and posterior insula to many subcortical and

cortical areas is coherent with integrative computations and high-level representations of

affective state:

AI is interconnected with subcortical structures such as the nuclei of the brain stem, limbic

structures and the basal ganglia, supporting its involvement in the representation and

integration of autonomic and visceral signals. Correspondingly, insula is also the cortical

terminus of visceral and ‘motivationally salient’ afferent fibres through spinal cord (Laminar

1) to brainstem and thalamus where spinal, humoral and vagus nerve interoceptive

information converge before projecting to posterior dorsal insular cortex. Somatosensory and

motor integration in posterior insula are supported by its connections to structures such as the

thalamus and basal ganglia.

Particularly important for our account are the AI’s dense connections with the nuclei of the

amygdala which is implicated in emotion (e.g., fear) and novelty detection.

In addition, AI connects bidirectionally to decision-making networks including areas

implicated in valuation and response selection such as OFC, nucleus accumbens, anterior

cingulate, dorsolateral prefrontal cortex (DLPFC, Brodmann area 46) and prefrontal areas.

Finally, intrainsular connections – required to represent and to compare predictive and actual

states – exist bi-directionally with stronger connectivity from AI to PI51-54.

19

Box 2: Interoceptive models of emotions

In the late 19th century, James and Lange suggested that the experience of emotions

necessarily depends on changes in bodily responses that automatically accompany emotive

stimuli20. Thus, we feel love when our hearts beat faster upon seeing our beloved and we feel

anxious when our stomachs constrict in response to the stress of making a difficult decision.

Although patterns of bodily responses can be linked to discrete emotions (e.g., our face

reddening with rage, blushing with embarrassment, paling with fear), bodily arousal states

remain rather undifferentiated. To better account for the variety of human emotional

experience while acknowledging a bodily contribution to feelings, Schachter and Singer23

refined this peripheral theory of emotion: They proposed a two-step model of emotions

according to which the onset and intensity of an emotion is determined by bodily signals and

the quality and category of the emotion is determined by one’s cognitive appraisal of the

context in which the change in bodily arousal was experienced (see also Klaus Scherer’s55

cognitive appraisal model). In recent years, these models have been rediscovered and

extended by neuroscientists, notably Damasio and colleagues (e.g., Ref. 21) in their

formulation of the somatic marker hypothesis. Here, body signals are seen as the necessary

basis in a hierarchy of conscious and emotional experiences, representing the core self in

terms of “I feel my body therefore I am” and also as a crucial influence on motivational

behaviour especially when one is making decisions involving uncertainty (see below). Bodily

arousal responses (‘somatic markers’ generated in response to the risk of a negative outcome)

are fed back to guide adaptive behaviour pre-consciously or in the form of ‘gut feelings’ and

‘hunches’. Patients with damage to medial prefrontal and somatosensory cortices manifest

emotional and behavioural difficulties consistent with acquired deficits in the generation and

feedback of somatic markers. An anatomically-inspired account proposed by Bud Craig

20

suggests that insular cortices support the cortical representation of interoceptive information

and awareness1.

Box 3: Uncertainty and risk

Uncertainty and risk

‘Uncertainty’ describes the degree to which past, present or future states and events are

known. It is induced by a lack of information (e.g., about the availability of rewards or about

how a sensory stimulus will feel to me). Imperfect measurements (e.g., due to the unreliability

of our sensory and visceral organs), poor predictive models (e.g., those employed when one

tries to predict a stranger’s behaviour in an ultimatum game) and random events (e.g., the

outcome of a coin toss) can all contribute to uncertainty.

Economics may further dissociate risk from uncertainty. ‘Risk’ describes the degree to which

all situational outcomes as well as their mathematical probabilities are known. For example, a

(fair) coin toss has 2 outcomes, each of which occurs with probability ½. Here, risk can be

modelled precisely, for example, as the variance of the outcome. In principle, (true or

estimated) probabilities can be assigned to any outcome, so the ideas presented above can be

applied to both risk and uncertainty, whether they are different concepts or two sides of the

same coin.

Learning about risk, uncertainty and feelings

One can learn about risk, uncertainty and feelings the same way one learns about rewards –

through reinforcement learning37. Comparing the risk prediction of an upcoming outcome

with the risk (or size of the error) measured at the time of the outcome results in risk

prediction error, which can serve to improve future risk predictions if taken into account.

21

Incorporating uncertainty into reinforcement learning (RL) algorithms

We can use an uncertainty measure to further improve predictions as well as prediction

learning about events (e.g., rewards) and states (e.g., feeling states). Imagine two predictive

cues, C1 and C2, where C1 reliably predicts stimulus strength S1 (low uncertainty) and C2

unreliably predicts stimulus strength S2 (high uncertainty). If you see both C1 and C2, what

stimulus strength should you expect? The best prediction (with minimal prediction error) is

one that puts more weight on C1 than on C2. On average, this will yield smaller prediction

errors. This reasoning extends to prediction learning. Predictions made in an uncertain

environment are likely to result in more frequent and larger errors. However, such errors are

less surprising (given that I was uncertain) than errors made in a certain environment. For

instance, an error after seeing C2 (high uncertainty) will be less surprising than an error after

seeing C1 (low uncertainty). As a result, I may increase my learning rate in response to the C1

error, but not the C2 error, because I knew that C2 was unreliable.

How uncertainty and uncertainty errors are tracked in the human brain and how uncertainty an

be incorporated into RL can be read elsewhere (e.g., Refs. 33,35,38).

Question Box: Questions for future research

- Can we spatially dissociate the representations of different states (modality-

specific feeling state, uncertainty, dominant integrated feeling state) which insula

is proposed to be involved in?

- Can we further identify different sites in AI that represent current and predictive

feeling states as well as feeling prediction errors?

- Are representations of current feeling states temporally correlated with

representations of predictive feeling states and feeling state prediction errors and

does this correlation depend on the uncertainty about the predictive feeling state?

22

- How are different modalities (e.g., feeling pain, disgust, pleasant tastes)

distinguished within insular cortex?

- Does insula represent feeling intensity or valence? How are positive and negative

valuation (feeling good or bad) distinguished within insular cortex?

- How is information about uncertainty integrated when one empathizes with and

vicariously learns about the feelings of others?

- Do functions of the insula influence decision making directly?

- How does insula function relate to memories about how things felt in the past? Is

this the same as a prediction of how things will feel in the future?

23

Figure caption

Figure 1: Anterior insula (AI) activation from fMRI studies on feelings, empathy and risk.

Panel A illustrates brain activity that is significantly correlated with changes in peripheral

electrodermal activity observed in subjects performing a gambling task56. Panel B depicts

enhanced bilateral anterior insula (AI) and right frontal opercular (FO) activity observed when

subjects performed an interoceptive task (heartbeat monitoring) versus an exteroceptive (note

monitoring) task22. Panel C illustrates bilateral AI activity when subjects empathized with

another person feeling pain. Activation shown reflects the average activation (N=78) observed

in three independent studies on empathy for pain 10,27,28. Panel D illustrates brain activity in

bilateral AI that is correlated with risk prediction (blue) and risk prediction errors (orange)

during a gambling task33.

Figure 2: A Model for the Integration of Feelings, Empathy and Uncertainty in Insular Cortex A. Schematic of error-based learning: A predicted state is followed by the actual

(experienced) state. The difference between the two, the prediction error, is used to update the

predicted state such that future predicted states will be more accurate. In the case of pain, the

predicted state is a predictive feeling state that is followed by the actual feeling state in

response to a painful stimulus. The prediction error with respect to the feeling indicates how

(in)accurate the predictive feeling state was. In the case of uncertainty, the predicted state is

the prediction risk that indicates how accurate I expect my prediction to be. The prediction

risk error that is generated at the outcome is used to update future estimates of prediction risk.

B. The integrated subjective feeling state combines information about modality-specific

feelings, uncertainty, contextual appraisal and individual preferences and traits such as risk

24

preferences and anxiety. No particular computational model is implied as to how the different

inputs are combined.

Figure Box 1. Illustration of the insula cortex in the human brain.

Figure 1

Figure 2

Figure Box 1