Interoceptive inference, emotion, and the embodied self

The concept of the brain as a prediction machine has enjoyed a resurgence in the context of the Bayesian brain and predictive coding approaches within cognitive science. To date, this perspective has been applied primarily to exteroceptive perception (e

The concept of the brain as a prediction machine has enjoyed a resurgence in the context of the Bayesian brain and predictive coding approaches within cognitive science. To date, this perspective has been applied primarily to exteroceptive perception (e.g., vision, audition), and action. Here, I describe a predictive, inferential perspective on interoception: 'interoceptive inference' conceives of subjective feeling states (emotions) as arising from actively-inferred generative (predictive) models of the causes of interoceptive afferents. The model generalizes 'appraisal' theories that view emotions as emerging from cognitive evaluations of physiological changes, and it sheds new light on the neurocognitive mechanisms that underlie the experience of body ownership and conscious selfhood in health and in neuropsychiatric illness.
The predictive brain, body, and self The view that prediction and error correction provide fundamental principles for understanding brain operation is gaining increasing traction within the cognitive and brain sciences. In the renascent guise of 'predictive coding' (PC -see Glossary) or 'predictive processing', perceptual content is seen as resulting from probabilistic, knowledgedriven inference on the external causes of sensory signals [1][2][3][4]. Here, this framework is applied to interoception, the sense of the internal physiological condition of the body [5], in order to elaborate a model of emotion as 'interoceptive inference' [6][7][8][9]. Interoceptive predictive coding -equivalently here, interoceptive inference -is hypothesised to engage an extended autonomic neural substrate with emphasis on the anterior insular cortex (AIC) as a comparator. This view extends alternative frameworks for understanding emotion [10][11][12][13][14] by proposing that emotional content is generated by active 'top-down' inference of the causes of interoceptive signals in a predictive coding context. It also extends previous models of insular cortex as supporting error-based learning of feeling states and uncertainty [15] and as responding to interoceptive mismatches that underlie anxiety [16].
Representations of physiological conditions have frequently been associated with basic pre-reflective forms of selfhood [11], with the AIC occupying a central role on some views [13]. Selfhood is a constellation concept that involves not only representation and control of physiological homeostasis, but also the experience of owning and identifying with a particular body, the emergence of a firstperson perspective, intention and agency, and metacognitive aspects that relate to the subjective 'I' and the narrative linking of episodic memories over time [17,18]. Here, I apply the framework of interoceptive inference to the experience of body ownership (EBO) as a central aspect of selfhood, proposing on the basis of recent data [19,20] that EBO is shaped by predictive multisensory integration Opinion Glossary Active inference: an extension of PC (and part of the free energy principle), which says that agents can suppress prediction errors by performing actions to bring about sensory states in line with predictions.
Augmented reality: a technique in which virtual images can be combined with real-world real-time visual input to create hybrid perceptual scenes that are usually presented to a subject via a head-mounted display. Appraisal theories of emotion: a long-standing tradition, dating back to James (but not Lange), according to which emotions depend on cognitive interpretations of physiological changes. Emotion: an affective state with psychological, experiential, behavioral, and visceral components. Emotional awareness refers to conscious awareness of an emotional state. Experience of body ownership (EBO): the experience of certain parts of the world as belonging to one's body. EBO can be distinguished into that related to body parts (e.g., a hand) and a global sense of identification with a whole body. Free energy principle: a generalization of PC according to which organisms minimize an upper bound on the entropy of sensory signals (the free energy). Under specific assumptions, free energy translates to prediction error. Generative model: a probabilistic model that links (hidden) causes and data, usually specified in terms of likelihoods (of observing some data given their causes) and priors (on these causes). Generative models can be used to generate inputs in the absence of external stimulation. Interoception: the sense of the internal physiological condition of the body. Interoceptive sensitivity: a characterological trait that reflects individual sensitivity to interoceptive signals, usually operationalized via heartbeat detection tasks. Predictive coding (PC): a data processing strategy whereby signals are represented by generative models. PC is typically implemented by functional architectures in which top-down signals convey predictions and bottom-up signals convey prediction errors. Rubber hand illusion (RHI): a classic experiment in which the experience of body ownership is manipulated via perceptual correlations such that a fake (i.e., rubber) hand is experienced as part of a subject's body. Selfhood: the experience of being a distinct, holistic entity, capable of global self-control and attention, possessing a body and a location in space and time [64]. Selfhood operates on multiple levels -from basic physiological representations to metacognitive and narrative aspects. Subjective feeling states: consciously experienced emotional states that underlie emotional awareness. Von Economo neurons (VENs): long-range projection neurons found selectively in hominid primates and certain other species. VENs are found preferentially in the AIC and ACC.
of self-related signals across interoceptive and exteroceptive domains. Overall, the model described here provides a unified view of self-related processing relevant to emotional awareness and EBO, and carries implications for understanding specific neuropsychiatric disorders.

Predictive inference and perception
The concept of PC overturns classical notions of perception as a largely 'bottom-up' process of evidence accumulation or feature detection, proposing instead that perceptual content is specified by top-down predictive signals that emerge from hierarchically organized generative models of the causes of sensory signals. According to PC, the brain is continuously attempting to minimize the discrepancy or 'prediction error' between its inputs and its emerging models of the causes of these inputs via neural computations approximating Bayesian inference ( Figure 1 and Box 1). Importantly, prediction errors can be minimized either by updating generative models (perceptual inference and learning; changing the model to fit the world) or by performing actions to bring about sensory states in line with predictions (active inference; changing the world to fit the model). In most incarnations these processes are assumed to unfold continuously and simultaneously, underlining a deep continuity between perception and Box 1. PC, free energy, and active inference PC has a long history, originating with the insights of von Helmholtz and reaching recent prominence in the 'Bayesian brain' hypothesis [1,4]. The idea is that, in order to support adaptive responses, the brain must discover information about the likely causes of sensory signals (i.e., perception) without direct access to these causes, using only information in the flux of sensory signals themselves [2]. According to PC, this is accomplished via probabilistic inference on the causes of sensory signals, computed according to Bayesian principles. This means estimating the probable causes of data (the posterior) given observed conditional probabilities (likelihoods) and prior 'beliefs' about probable causes. This, in turn, means inducing a predictive or 'generative' model of the sensory data. Although exact Bayesian inference is computationally challenging and often intractable, a variety of approximate methods exist. Within neuroscience, these approximations have been elaborated in Friston's 'free energy principle' [2,65], which, following seminal work by Hinton and colleagues [66,67], shows how generative models can be induced from data by assuming that the brain minimizes a bound on the evidence for this data (the 'free energy', which under simplifying (Gaussian) assumptions is equivalent to prediction error). The generalization of Bayes theorem to a hierarchical context implies that posteriors at one level form the priors at one level lower, thus enabling priors to be induced from the data stream itself ('empirical' Bayes). Applied to cortical networks, PC interprets bottom-up signals as conveying prediction errors and top-down signals as conveying predictions ( Figure 1). Although unequivocal neural evidence for PC is still lacking, a growing body of supportive data details how perceptual content -and underlying neural responses -can be shaped by prestimulus expectations [1,4,42]. Key to PC is the minimization of prediction error across hierarchical levels. This can be accomplished either by updating generative models to accommodate unexpected sensory signals or by performing actions to confirm sensory predictions (active inference, [21]). This duality underlines a strong continuity between perception, action, and imagination [68]. Also important is that prediction errors are associated with precisions ( Figure 1), so that dynamic precisionweighting (for example by attention) can modulate the balance between top-down and bottom-up signal flow (e.g., low precision on error signals corresponds to high confidence in top-down prior beliefs). The present framework generalizes PC to interoception, proposing that affective states depend on active inference of interoceptive responses.

Opinion
Trends in Cognitive Sciences November 2013, Vol. 17,No. 11 action. Prediction errors are associated with 'precisions', which determine their influence on subsequent hierarchical processing. For example, precision weighting (possibly implemented by post-synaptic gain modulation of prediction error units) can modulate the extent to which prediction errors are resolved by updating generative models or by performing actions [21]. This leads to an interpretation of attention as the optimization of precision weighting, balancing the relative influence of prediction errors and prior expectations on perceptual inference [2].
PC has been elaborated principally in the context of exteroception; that is, to predictive modelling of external states of the world. However, one of the most relevant features of the world for a particular organism is the organism itself [11,22]. This reflects a long-standing notion that mental representations of selfhood are ultimately grounded in representations of the body, with the internal physiological milieu providing a primary reference -a 'material me' [13] -that supports adaptive interaction with the environment. From a PC perspective, this implies that an organism should maintain well-adapted predictive models of its own physical body (its position, morphology, etc.) and of its internal physiological condition. This entails inducing generative models of the causes of those signals 'most likely to be me' [22] across interoceptive and exteroceptive domains, a framework that views emotion as 'interoceptive inference' and provides a unifying mechanism for self-representation at multiple levels, including perhaps especially those related to EBO.

Interoceptive inference and emotion
Interoceptive concepts of emotion were crystallized by James and Lange [10], who argued that emotions arise from perceptions of changes in the body. This approach evolved into 'appraisal' theories, which recognise that explicit cognitions and beliefs about the causes of physiological changes influence subjective feeling states and emotional behaviour [23]. Schachter and Singer [24] famously demonstrated that injections of adrenaline, proximally causing a state of physiological arousal, would give rise to either anger or elation depending on the context (an irritated or elated confederate). This observation was formalized in their 'two factor' theory, in which emotional experience is determined by the combination of physiological change and cognitive appraisal, that is, emotion as interpreted bodily arousal (see [25] for a precursor). More than a century after James and Lange, there is now a consensus that emotions are psychological states that encompass behavioural, experiential, and visceral changes [23,[26][27][28]. This attitude underpins several contemporary frameworks for understanding emotion and its relation to cognition and self [11][12][13], discussed further below.
Despite the above insights, interoception has remained generally understood along feed-forward lines, similar to classical evidence accumulation theories of exteroception [23,28]. This assumption is however challenged by evidence of substantial cross-talk between levels of viscerosensory representation, including top-down cortical and behavioural influences to brainstem and spinal centres [26]. Informed by this emerging picture, I suggest that the role of interoception in shaping emotion and selfhood can be productively understood through the lens of PC. In this view, interoceptive inference involves hierarchically cascading top-down interoceptive predictions that counterflow with bottom-up interoceptive prediction errors. Subjective feeling states -experienced emotions -are hypothesized to depend on the integrated content of these predictive representations across multiple levels [6].
Following PC principles, interoceptive prediction errors can be suppressed both by modifying predictions and by transcribing these predictions into reference points for autonomic reflexes that regulate physiological homeostasis, as recently suggested by Gu and colleagues [9]. This role for active inference, which extends previous presentations of this model [6][7][8], directly parallels PC formulations for motor control (e.g. [29]) which highlight descending corticospinal signals as instantiating proprioceptive predictions that engage classical motor reflexes. Precisions play a key role here: descending predictions can engage motor or autonomic reflexes only if the corresponding error signals have diminished impact on hierarchical processing via transiently low precision weighting, which corresponds to decreased attention to these error signals. Without this transient modulation, precise prediction errors would lead to revision of predictions rather than to action [29]. This implies that active interoceptive inference depends on the selective attenuation of attention to interoceptive prediction errors.
Interoceptive predictions arise from multiple hierarchical levels, with higher levels integrating interoceptive, proprioceptive and exteroceptive cues in formulating descending predictions. These multimodal predictions underwrite emotional responses to exteroceptive cues (which may include socially salient signals, see later). In short, interoceptive predictive coding (inference) proposes that emotional content is determined by active inference on the likely internal and external causes of changes in the physiological condition of the body ( Figure 2).

Evidence for interoceptive inference
Although there is not yet any direct confirmatory evidence for interoceptive inference (as for PC generally, see [1]), supportive data are steadily accumulating. Much of these data rest on assuming a central role for the anterior insular cortex (AIC), operating within a rich functional network [30], both as a comparator that registers top-down predictions against bottom-up prediction errors and as a source of anticipatory visceromotor control [6][7][8][9]15,16]. Structurally, the AIC is ideally placed both to detect and to cause changes in physiological condition, and to integrate interoceptive and exteroceptive signals; functionally, it instantiates interoceptive representations accessible to conscious awareness and is associated with processes that involve visceral representation, interoception, and emotional awareness relevant to selfhood (Box 2).
Several functional MRI (fMRI) studies have shown anticipatory and prediction error responses within the AIC. For example, the AIC is activated by anticipation of painful [31] and affect-laden touch [32], and AIC responses encode both predicted pain and pain prediction error within a single task [33]. Similarly, in a gambling task, the AIC has been shown to encode both predicted risk Opinion Trends in Cognitive Sciences November 2013, Vol. 17,No. 11 and risk prediction error [34]. Holle et al. found AIC activation on viewing movies of people scratching, with the degree of activation correlating with the subjective level of reported itchiness [35]. Here, AIC responses plausibly reflect a conflict between predicted and actual interoceptive responses given anticipated itch, which suggests an engagement of AIC in predictive inference related to mirror-system activity and empathy [36]. Using the neuroeconomic 'ultimatum game', Xiang and colleagues showed that subjects used Bayesian rules to update expected feelings given economic outcomes, and that feeling prediction errors (when measured using variance) scaled parametrically with AIC activation [37]. These studies support the idea that interoceptive inference extends into social contexts to explain emotional behaviour. For example, the expression of an emotional state (e.g., anger) elicits behavioural responses from others, the detection of which could serve to confirm predictions of interoceptive condition. This provides a nice link to predictive models of social interaction [38,39].
Supportive evidence is also provided by studies that manipulate interoceptive feedback. Early evidence showed that false cardiac feedback enhanced subjective ratings of emotive stimuli [40]. This was later confirmed by Gray and colleagues, who found that false cardiac feedback led to increased fMRI responses in the right AIC and that enhanced AIC activity during false feedback correlated with increased emotional salience attributed to previously unthreatening stimuli [41]. Further evidence showing the influence of false interoceptive feedback on the experience of body ownership is described below. Taken together with the structural and functional attributes of the AIC (Box 2),  Interoceptive predictions are proposed to be generated, compared, and updated within a salience network (orange shading) anchored on the anterior insular and anterior cingulate cortices (AIC, ACC) that engage brainstem regions as targets for visceromotor control and relays of afferent interoceptive signals. Sympathetic and parasympathetic outflow from the AIC and ACC are in the form of interoceptive predictions that enslave autonomic reflexes (e.g., heart/respiratory rate, smooth muscle behaviour), just as proprioceptive predictions enslave classical motor reflexes in PC formulations of motor control [9]. This process depends on the transient attenuation of the precision of interoceptive (and proprioceptive) PE signals. Blue (red) arrows signify top-down (bottom-up) connections respectively.

Box 2. The anterior insula cortex and interoception
The human insular cortex is found bilaterally beneath the temporal and frontal lobes, enjoying widespread bidirectional connectivity to parietal, frontal, and limbic regions [30]. Interoceptive pathways have their primary cortical representation within the insula, which contains a viscerotopic map [69]. A posterior-to-anterior gradient has been proposed, whereby posterior regions support primary (objective) mappings of interoceptive signals, whereas the anterior insula (AIC) supports secondary re-representations that underlie subjective access [9,12,13] and integration of interoceptive, motivational, and exteroceptive signals [6,70]. The AIC is closely connected structurally and functionally with the anterior cingulate cortex (ACC) as part of a cortical 'salience network' [71], though they can be functionally differentiated [70], so that the AIC can be considered a 'limbic sensory area' and the ACC a 'limbic motor area' [13,70]. These areas together engage subcortical regions, such as the periaqueductal gray matter and parabrachial nucleus, as targets for visceromotor control and relays for viscerosensory afferents, as well as many other areas related to self and emotion, including the amygdala, nucleus accumbens, and orbitofrontal cortex. Interestingly, the AIC contains a high density of von Economo neurons (VENs) [72], which have been associated frequently though circumstantially with conscious awareness and selfhood [8,13,73]. VENs are large projection neurons well suited for rapid long-range information integration and so may underlie the efficient registration of interoceptive prediction and prediction error signals. In PC, topdown predictions are suggested to originate in infragranular pyramidal cells (Figure 1), consistent with the layer V predominance of VENs in the AIC [8,9,73]. The AIC is engaged in a wide range of processes that share as common factors visceral representation, interoception, and emotional awareness [6,9,13,26]. Importantly, the AIC (particularly on the right) appears to support conscious access to both interoceptive information -as reflected in subjective emotional states (though see [43,74])and to associated representations of how encoded objects and contexts relate to the biological self [15,75]. Distinct axes of interoceptive information map onto specific insula sub-regions giving rise to dissociable emotional feeling states [76]. Also, individual differences in interoceptive sensitivity (IS), as measured by heartbeat detection, are predicted by AIC activation and morphometry, and are also associated with reported emotional symptoms [12] and susceptibility to illusions of selfhood (see main text).
These structural and functional attributes, when considered together with evidence for prediction error processing, strongly suggest the AIC as a locus for comparator mechanisms that underlie interoceptive inference, in turn supporting emotional awareness and consciously accessible integrated self-representations.

Opinion
Trends in Cognitive Sciences November 2013, Vol. 17,No. 11 these data provide indirect support for interoceptive inference and the central involvement of the AIC in prediction error registration underlying subjective feeling states. Additional experiments, perhaps based on designs utilized in exteroceptive contexts (e.g., [42]), are needed to test more directly interoceptive inference and the proposed role of the AIC.

Relation to other models of emotion and insula function
The present model is related to several contemporary frameworks for understanding emotion and self, including prominently those of Damasio [11], Craig [13], and Critchley [12]. Each of these sees selfhood as grounded in representations of physiological condition, with the AIC emphasised in some [12,13], but not all [43]. Crucially, none identify emotional states with top-down inference of the causes of interoceptive signals, as argued here (and as recently taken up by others [9]). Rather, they emphasize a continuous, dynamic, but largely bottom-up interoceptive representational hierarchy that interacts with other perceptions to motivate behaviour. Perhaps more closely related is the suggestion that the AIC is involved in errorbased learning of feeling states and uncertainty [15]; here, the notion of 'prediction error' is expressed in terms of change detection and salience rather than through mechanisms of predictive coding. Also relevant are models of insular dysfunction during anxiety [16] and psychosis [44] that hypothesize the existence of interoceptive prediction errors. However, although these models integrate abundant evidence compatible with a role for predictive interoception underlying emotion and self, the core notion of interoceptive inference is not elaborated.
The predictive self and the experience of body ownership A predictive model of selfhood must extend beyond subjective feelings to integrate interoceptive and exteroceptive signals across multiple levels of self-representation. Particularly significant is the representation of those parts of the world perceived as belonging to one's own body, supporting EBO [45], and which plausibly are closely tied to interoception in having the body as a referent. Experiments such as the 'rubber hand illusion' (RHI) attest to the plasticity of EBO: stroking of an artificial hand synchronously with a participant's real hand, while visual attention is focused on the artificial hand, leads the participant to experience the artificial hand as part of her own body [46]. Similar effects have been described for face perception [47] and whole-body ownership [48,49].
These manipulations of EBO have been explained by models of multisensory integration, which propose that conflicts between vision, touch, and proprioception are minimized by visual capture of visual and felt (tactile) events occurring in peri-personal space, on the basis of statistical correlations among sensory signals together with visual dominance [46,50]. This is compatible with PC inasmuch as minimization of prediction errors -such as those induced by multisensory conflicts during the RHI -will update the posterior probabilities and, over time, can induce changes in self-related priors [22]. Priors reflecting high-level representations, such as selfhood and body-ownership, are likely to operate at relatively abstract multisensory or amodal levels. Thus, statistical correlations among highly precision-weighted sensory signals (vision, touch) could overcome prediction errors in a different modality (proprioception), leading to a revised multisensory predictive model that minimizes the overall level of self-related precision-weighted prediction error by incorporating the fake hand as part of the selfrepresentation. Moreover, the relative weighting of different information sources according to their reliability in this process is suggestive of the deployment of precision expectations which underlie optimal cue combination in a Bayesian scheme [51]. The engagement of predictive self-models is further supported by recent evidence showing that perceptual correlations are not necessary for the RHI; the expectation of correlated sensory input is sufficient [52]. Interestingly, EBO has also been associated with the integrity and activity of the AIC on the basis of lesion data and functional neuroimaging in healthy subjects [53].
Experiments have begun to address the impact of interoceptive processing on the modulation of EBO in paradigms like the RHI. One avenue is provided by individual differences in 'interoceptive sensitivity' (IS), which refers to a person's ability to detect their own interoceptive signals, operationalized by heartbeat detection tasks [54]. Tsakiris and colleagues reported that participants with lower IS are more susceptible to the RHI [55] and to the modulation of self-other boundaries in response to multisensory stimulation [56,57], possibly reflecting lower precision-weighting of interoceptive prediction errors. In addition, induction of the RHI leads to decreased temperature [58] and increased histamine reactivity of the real hand [59], whereas cooling the real hand increases susceptibility to the RHI [60]. Also, threats to the rubber hand during the RHI evoke enhanced skin conductance responses [61]. These effects are compatible with descending self-related predictions tuning reference points for autonomic reflexes from the perspective of active inference.
Two recent studies have addressed directly the role of multisensory integration across interoceptive and exteroceptive domains in shaping EBO. In the first, Suzuki et al. [19] combined augmented reality with physiological monitoring to implement a version of the RHI in which a virtual hand changed colour either in-time or out-of-time with an individual's heartbeat (Figure 3). The authors found that synchronous cardio-visual feedback enhanced the experience of ownership of the virtual hand, as measured both subjectively by questionnaire and objectively by previously validated measures of 'proprioceptive drift' -where the latter measures the perceived position of the real (hidden) hand by asking the participant to move a cursor to this location. (Note that these two measures do not always correlate [62].) The same setup enabled a replication of the standard RHI by contrasting synchronous versus asynchronous tactile-visual feedback. In a third condition, Suzuki et al. found that real-time remapping of finger movements to the virtual hand -supporting peri-hand visual-proprioceptive coherence -provided strong cues for EBO that overshadowed the influence of cardiac (interoceptive) prediction errors. A second study [20] Opinion Trends in Cognitive Sciences November 2013, Vol. 17,No. 11 applied similar cardio-visual feedback to a virtual body as seen from behind, thus testing the influence of such feedback on the 'full body illusion' [48,49]. Consistent with Suzuki et al.'s results, they found that cardio-visual synchrony increased identification with the virtual body, again when measured both objectively and subjectively.
These data suggest that statistical correlations between interoceptive (e.g., cardiac) and exteroceptive (e.g., visual) signals can lead to updating of predictive models of selfrelated signals through minimization of prediction error, just as may happen for purely exteroceptive multisensory conflicts in the classic RHI. Also, if predictive models are continually probed by control signals that attempt to confirm the currently dominant model via active inference [21], the framework naturally accommodates the phasic physiological changes that accompany RHI induction [58,59] when these changes are taken to reflect altered autonomic control.

Concluding remarks
This opinion article proposes that emotion and embodied selfhood are grounded in active inference of those signals most likely to be 'me' across interoceptive and exteroceptive domains. In humans, self-related predictive coding simultaneously engages multiple levels of self-representation, including physiological homeostasis, physical bodily integrity, morphology and position, and -more speculatively -the metacognitive and narrative 'I'. Subjective feeling states (emotional experiences) arise from active interoceptive inference, extending previous theories based on cognitive appraisal of perceived physiological changes [24] and contemporary frameworks that emphasize  Subjects wore a front-facing camera connected to the HMD, so they saw the camera image superimposed with the virtual hand. They also wore a pulse-oximeter to measure heartbeat timings and they used their right hand to make behavioural responses. (B) Cardio-visual feedback (left) was implemented by changing the colour of the virtual hand from its natural colour towards red and back, over 500 ms either synchronously or asynchronously with the heartbeat. Tactile feedback (middle) was given by a paintbrush, which was rendered into the AR environment. A 'proprioceptive drift' (PD) test (right), adapted for the AR environment, objectively measured perceived virtual hand position by implementing a virtual measure and cursor. (C) The experiment consisted of three blocks of four trials each. Each trial consisted of two PD tests flanking an induction period, during which either cardio-visual or tactile-visual feedback was provided (120 s). Each trial ended with a questionnaire presented in the HMD. (D) PD differences (PDD, postinduction minus pre-induction) were significantly larger for synchronous versus asynchronous cardio-visual feedback in the 'cardiac still' (without finger movements), but not the 'cardiac move' condition (with finger movements). PDDs were also significantly larger for synchronous versus asynchronous tactile-visual feedback ('tactile' condition), replicating the classical RHI. Each bar shows the across-participant average and standard error. (E) Subjective questionnaire responses that probed experience of ownership showed the same pattern as PDDs, whereas control questions showed no effect of cardio-visual or tactile-visual synchrony. Adapted from [19].

Opinion
Trends in Cognitive Sciences November 2013, Vol. 17,No. 11 bottom-up elaboration of interoceptive representations with perception and motivation [11,13,15]. The close interplay between interoceptive and exteroceptive inference implies that emotional responses are inevitably shaped by cognitive and exteroceptive context, and that perceptual scenes that evoke interoceptive predictions will always be affectively coloured. Although the detailed neuroanatomy that underlies interoceptive inference remains to be elucidated, accumulating evidence implicates the AIC as a key comparator mechanism sensitive to interoceptive prediction error signals, as informing visceromotor control, and as underpinning conscious access to emotional states (emotional awareness). A predictive self is further supported by emerging paradigms that combine virtual/augmented reality and physiological monitoring, where the data so far suggest that the experience of body ownership, a key aspect of selfhood, is modulated by predictive multisensory integration of precision-weighted interoceptive and exteroceptive signals.
This framework may have important implications for understanding psychiatric disturbances of selfhood and emotional awareness. Dysfunctions in interoceptive inference could underlie a range of pathologies, especially those that involve dissociative symptoms, such as anxiety (Box 3). A broad role for interoceptive inference is also consistent with recent evidence that shows interoceptive influences on cognition and perception. For example, word recognition memory is modulated by the timing of visual stimuli with respect to cardiac phase under restricted attention [63]. Overall, 'interoceptive inference' highlights common predictive mechanisms that underlie affect, self, and perception, emphasizes their integration in shaping self-related experience, and provides new experimental possibilities for elucidating the underlying processes (Box 4).

Box 4. Questions for future research
Can (structurally or functionally) segregated populations of representation and error units be identified in the interoceptive system? There are some early indications of such segregation in perceptual systems [88]. Do neurotransmitters like dopamine and oxytocin have a role in modulating the precision of interoceptive prediction errors [89]? How do these roles relate to similar roles in reward processing [2]? Are VENs engaged in communicating top-down interoceptive predictions from insular and cingulate cortices to subcortical and brainstem targets [8]? What are the functional and effective connectivity patterns that underlie interoceptive inference and the integration of interoceptive and exteroceptive signals relevant to conscious selfhood? What is the relation between predictive models of interoception and similar models of agency [90], another central aspect of selfhood? Can predictive models of selfhood accommodate more abstract modes of self-experience, such as the narrative self or 'I' that links episodic memories across time? Can interoceptive feedback provide new avenues for the treatment of emotional and neuropsychiatric disorders? Might there be a role for disrupted interoceptive inference in impairments of social emotional behaviour, such as autism? Box 3. Interoceptive inference and psychopathology A significant body of work now connects deficits in predictive inference with psychiatric disorders that affect self-representation [6,16,44]. For two decades, 'comparator' models of schizophrenia have suggested that disturbances of selfhood (e.g., delusions of control) reflect problems in distinguishing self-caused from externally-caused changes in sensory input [77]. In psychosis, false perceptions (hallucinations) and false beliefs (delusions) may arise from reshaping of top-down predictions in attempts to suppress aberrant and persistent (exteroceptive) prediction errors [78]. This view has been finessed in terms of abnormal encoding of the relative precision of priors and sensory evidence to account for a broad range of psychotic symptoms [79,80].
Considering interoceptive inference and the AIC as its likely brain basis further enhances the explanatory potential of this framework for psychopathology. Links between perceptions of bodily state and emotional and neuropsychiatric disorders are well established. For example, alexithymia (broadly, deficits in emotional awareness) is associated with failure to engage AIC [81] and with degeneration of AIC and of VENs (Box 2) in the context of fronto-temporal dementia [82,83]. Human VENs express proteins linked to schizophrenia (notably DISC-1) and VEN density (in the anterior cingulate) has been linked to illness duration and completed suicide in psychotic patients [84]. Neuroimaging studies have correlated (right) AIC activity and volume with individual interoceptive sensitivity (IS), which in turn predicts sensitivity to mood states, including anxiety [12,16]. Indeed, anxiety and psychosis have been specifically associated with mismatches between predicted and actual interoceptive states putatively computed within AIC [16,44]. The finding that low IS predicts susceptibility to the RHI [55] provides a bridge to understanding how disorders of body image and ownership may involve disrupted interoceptive inference.
Dissociative conditions, such as depersonalization and derealisation (DPD), involve disabling disturbances of selfhood, as reflected in a persistent sense of unreality ('as-if-ness') [85]. These conditions are associated with alexithymia and a general loss of 'emotional colour', suggestive of deficient interoceptive inference and consistent with observed hypoactivation of AIC (and ACC bilaterally) in DPD patients when viewing aversive images [86]. By analogy with models of psychosis [79,80], Seth et al. have suggested that DPD may arise from imprecise (as opposed to inaccurate) interoceptive predictions, as part of a model of conscious 'presence' [6]. The transition from DPD to full-blown delusion (e.g., Cotard's syndrome, in which patients believe they are dead [87]) may also reflect aberrant high-level inference as the result of attempting to explain away persistent interoceptive prediction errors.