Introduction

Far more complex than the mere transmission of nociceptive inputs, pain is a component of the interoceptive system (Craig, 2003). It is defined as a protective experience that can occur even in the absence of physical harm (Loeser & Treede, 2008) and is permeated by all the available sources of information, including sensory, emotional, cognitive, and social components as well as prior information and what one expects to happen (Atlas & Wager, 2012; Tracey, 2010; Williams & Craig, 2016). While the last decades provided important advances in the understanding of pain, developing an overarching theory that accounts for all the dimensions of pain still remains challenging (for a full discussion, see Moayedi and Davis, 2013).

Current views propose predictive processing models as a suitable candidate to account for the multifaceted experience of pain (Büchel et al., 2014; Kiverstein et al., 2022; Song et al., 2019, 2021). Interoceptive inference and the Embodied Predictive Coding frameworks (for a full discussion see Barrett, 2017; Barrett and Simmons, 2015; Pezzulo, 2014; Seth, 2013; Seth and Friston, 2016) conceptualise the brain as an active generator of inferences, which constantly updates an internal model of both the internal and the external world and attempts to fit it with incoming inputs through a process of Bayesian hypothesis testing and revision. Interoceptive sensations are thought to be derived by the integration of different sources of information into an expectation of the body’s upcoming changes, which are kept in check by the actual state of the body (Barrett & Simmons, 2015). However, people’s predictions do not always mirror the reality, but they can be often inaccurate, leading to a gap between the expected and the actual present state. The goal of inferential processes is to reduce the difference between one’s internal models and the actual interoceptive input, thus converging towards the brain’s best guess about the body’s state (Ainley et al., 2016; A. Seth et al., 2012; A. K. Seth & Friston, 2016).

An important feature of these accounts is that the difference between one’s internal models and the actual sensory data can be minimized not only by updating internal models to better fit the data (i.e., perceptual inference), but also by performing actions, changing bodily states themselves (i.e., active inference) (for a full discussion, see Friston, 2010, 2005; Parr et al., 2022). For example, an expectation of a threatening, painful stimulus could be realised both by making the painful stimulus appear more painful than it really is, thereby fitting one’s internal model to bodily reality, or by actually modulating the state of the body and shape the physiologic response to pain towards previous predictions, for example by engaging autonomic reflexes (e.g., typical decrease in the heart rate in relation to pain anticipation, e.g., Bradley et al., 2008, 2005; Colloca et al., 2006; Lykken et al., 1972; Taggart et al., 1976; Tracy et al., 2017).

There is suggestive evidence for both components in the extant literature on pain. The assumption that prior information modulates pain perception underlines decades of work on placebo and nocebo effects (i.e., the expectation-effect; Crombez et al., 1998; Keltner et al., 2006; Petrovic and Ingvar, 2002; Price et al., 1999; Wager, 2005; Wager et al., 2004). On a neuronal level, expecting pain alters the neural mechanisms in pain-processing regions, thought to reflect a combination of nociceptive inputs and top-down information (Atlas et al., 2010; Keltner et al., 2006; Koyama et al., 2005; Wiech et al., 2008), even at a very early stage of processing (Eippert et al., 2009; Geuter & Büchel, 2013). Moreover, such changes can persist – or even grow over time – in the absence of disconfirming evidence (Atlas et al., 2010; Colloca et al., 2010; Craggs et al., 2008; Jepma & Wager, 2015; Koban & Wager, 2016; Montgomery & Kirsch, 1997; Vase et al., 2005, 2011), turning pain experience into potentially a self-fulfilling prophecy (Jepma et al., 2018).

As for the assumption that unfulfilled predictions can be resolved by changing the state of the body and engaging autonomic reactions, there is evidence that prior pain expectations, and pain itself, modulate physiological responses. Expecting pain induces augmented activity in the sympathetic system (e.g. increases in blood pressure and skin conductance) to prepare potential avoidance responses (Barlow et al., 1996; Oka et al., 2007; Tousignant-Laflamme & Marchand, 2006; Yang et al., 2003), as well as a characteristic decrease in heart rate to promote orienting responses, attention and sensory processing (Bradley et al., 2008, 2005; Colloca et al., 2006; Lykken et al., 1972; Taggart et al., 1976; Tracy et al., 2017; for a full discussion, see Skora et al., 2022)

Collectively, these findings support the idea that the experience of pain relies on top-down predictive mechanisms, which affect not only the subjective perception of pain, but also its neural and bodily correlates. However, in previous work pain expectations were typically induced by exteroceptive cues (e.g., visual, Wiech et al., 2014; Jepma et al., 2019; auditory, Colloca et al., 2006; Atlas et al., 2010), neglecting the role of interoceptive sources of information within the inference. Among such streams, the activity of the heart appears to be tightly linked with the experience of pain, probably due to a close integration of the pain system and the neural network involved in cardiovascular regulation ((Bud) Craig, 2003; Craig, 2003, 2008). For example, pain-related evoked potentials and nociception can be modulated over the cardiac cycle (Edwards et al., 2001, 2002, 2008; Martins et al., 2009; McIntyre et al., 2006), and either the anticipation or the response to pain stimuli is associated with changes in heart rates (Colloca et al., 2006; Mischkowski et al., 2018; for a review, see Kyle and McNeil, 2014). Moreover, when estimating their cardiac frequency, people’s predictions draw upon the level of threat (i.e., pain) associated with upcoming events (Parrotta, 2022), reflecting the (fictitious) belief that threat expectations lead to an increase in heart rates. Even though anticipating pain typically goes along with decreases in heart rates (the well-known orienting response Bradley et al., 2008, 2005; Colloca et al., 2006; Lykken et al., 1972; Taggart et al., 1976; Tracy et al., 2017), people may therefore illusorily perceive a higher cardiac frequency when expecting painful stimulation (Parrotta et al., 2022).

The tight coupling between heart rates changes and pain experience suggests that it should be possible to modulate the experience of pain by manipulating interoceptive cardiac feedback, that is, by misleading people into believing that their heartrate is increasing in anticipation of a noxious stimulus. Under predictive architectures, persuading one’s internal model towards the fictitious evidence of increased heart rates (i.e., faster cardiac feedback manipulation) would potentially induce consequent changes both in pain perception (i.e., perceptual inference) and in the actual state of the body (i.e., active inference). Specifically, the former process would correspond to an interoceptive illusion of increased pain (where “illusion” is used in the same sense as in bodily illusions, e.g., Ferri et al., 2013; Tsakiris et al., 2011), whereas the latter would correspond to physiological and autonomic adaptations to the predicted level of pain. While it has been consistently reported that exteroceptive cues induce predictions about pain, it is currently unknown whether a cardiac feedback manipulation that renders interoceptive streams more similar to those expected in pain conditions would produce an interoceptive illusion of pain and ensuing autonomic adjustments - as hypothesized by interoceptive inference and embodied predictive coding theories. To test this hypothesis, Experiment 1 provided participants with auditory cardiac feedback, via headphones, that was either congruent with their heartbeat frequency, or incongruent with it (i.e., faster or slower) before the administration of a noxious electrical stimulation, while their ECG was recorded. Once the pain stimulus was administered, subjects were asked to rate the intensity and unpleasantness of the painful stimuli.

We predicted, first, that the exposure to faster vs. slower cardiac feedback induces expectations of increased levels of pain, such that noxious stimuli are felt and reported as more unpleasant and intense. Second, expectations of threat induced by the faster feedback should be fulfilled through parasympathetic reflexes, such that participants’ real heartrates decrease over the exposure to the incongruent faster, compared to the slower cardiac feedback, in line with the well-known orienting cardiac response when expecting threat or pain (Bradley et al., 2008, 2005; Colloca et al., 2006; Lykken et al., 1972; Taggart et al., 1976; Tracy et al., 2017; for a full discussion, see Skora et al., 2022). To rule out that potential modulations in participants’ pain perception and cardiac state were merely associated with the frequency of the feedback, rather than an heartbeat, Experiment 2 replicated the design in a second group of participants, with the only difference being that the congruent or incongruent feedback would not be a heartbeat tone, but an unrelated exteroceptive stimulus. We predicted that as participants would have no expectation that such an external stimulus signals a preparation for a threatening, painful stimulus, variations in the rate of this exteroceptive stimulus should lead to no (or less) perceptual and cardiac changes.

If confirmed, the results would provide the first evidence that predictive processes are generated by the simulation of interoceptive streams (i.e., cardiac feedback manipulation), influencing not only our perception of pain but also its heart-related physiological response: two changes that jointlyminimize the discrepancy between the expected and the actual homeostatic (interoceptive) state. This evidence would corroborate the assumption that predictions originate along the embodied multisensory sphere, by orchestrating multiple sources of information (i.e., interoceptive) within the inference. Ultimately, the results may provide important insights into how the experience of pain is actively built in light of predictive processes, shedding new light into the understanding of pain-related pathologies in terms of chronic aberrant predictions, as already hypothesized for other dysfunctional conditions (Corlett et al., 2019; Gagne et al., 2018; Lissek & van Meurs, 2015; Powers et al., 2017; Series, 2019), laying the groundwork for new avenues for the future comprehension of pain and its treatment.

Experiment 1

Experiment 1 tests whether false cardiac feedback of an accelerated (i.e., faster), relative to a decelerated (i.e., slower) heartrate alters both pain perception and the cardiac anticipatory response to noxious stimuli.

Methods

Participants

A sensitivity analysis with G*Power 3.1 (Faul et al., 2007) showed that a sample size of 34 provides .90 power to detect effects with Cohen’s d =0.57 (SESOI of δ = 0.34).Thirty-four participants (mean age 25.11, SD= 2.94, 21 women) took part in the experiment, recruited from Gabriele D’Annunzio University and the wider community. All were right-handed with normal or corrected-to-normal vision. Exclusion criteria for taking part were self-reported chronic and acute pain, neurological disease, serious cardiovascular disease (i.e., any type of disease involving the heart or blood vessels that might result in life-threatening medical emergencies, e.g., arrhythmias, infarct, stroke), or conditions that could potentially interfere with pain sensitivity (e.g. drug intake or skin diseases). Participants were asked to either not drink coffee or smoke cigarettes in the 60 minutes preceding the experiment. All gave written informed consent, were unaware of the purposes of the study, and were fully debriefed about it at the end of the experiment. Ethical approval from the local ethics board was obtained. One participant was excluded as, at the end of the experiment, they declared to suffer from a cardiac defect that made them realize the manipulation of the cardiac feedback.

Apparatus

Painful stimuli were electrical pulses delivered using a constant-current electrical stimulators (Digitimer DS7A) controlling a pair of neurological electrodes attached on the phalanx of the middle finger of the participant’s left hand, which provide a precise constant current, isolated stimulus, controllable in pulse duration and amplitude. The intensity of the electrical stimuli (2 ms) was kept fixed over the experiment and it was established during a calibration phase before the experiment.

Cardiac recording was performed by a Biopac MP 160 system (Biopac Systems Inc., USA). ECG was recorded continuously from two electrodes attached to the lower ribs and one over the right mid-clavicle bone (reference electrode). The ECG signal was sampled at 2 kHz with the Biopac Acqknowledge 3.7.1 software (Biopac Systems Inc., USA) according to the manufacturer guidelines. ECG signal was then fully analysed in MATLAB (R2020a). The tone used for the creation of the feedback was the sound of a single heartbeat, gathered from https://freesound.org and manipulated in Audacity. The feedback audio was then created in MATLAB(R2020a), repeating the single heartbeat sound according to the desired frequency (see Procedure).

Stimulus presentation was controlled through E-Prime (Psychology Software Tools Inc., Pittsburgh, USA), which interfaced with the pain stimulators and the Biopac system via a parallel port.

Procedure

Upon arrival at the lab, participants were briefed by the experimenter. After providing consent, theywere placed in a comfortable chair, and the ECG electrodes were applied after cleaning the skin. To increase the ambiguity, and thus predictive influences acting on the pain stimulus (Yon & Frith, 2021), the intensity of the noxious input varied between five intensities, which were identified for each participant in an initial calibration session. To do so, participants were informed that they would undergo a psychophysical calibration procedure to determine their subjective response to increasing stimulus intensities. They were comfortably seated with their left hand placed on a table. Before electrodes were placed, the skin was cleaned with alcohol to reduce impedance. The first stimulus was delivered at a low intensity, which is below the threshold for pain perception in most people. The intensity increased in a ramping procedure up to a maximum of five volts. Participants verballyrated the pain intensity for each stimulus using a 0-100 Numerical Pain Scale (NPS).

Following previous research (Atlas et al., 2014; Colloca et al., 2006; Hird et al., 2019) a pain intensity rating of NPS 20 denoted “just painful”, NPS 50 denoted “medium pain”, and NPS 80 marked the point at which the stimulus was “just tolerable”. We identified a ‘low’ pain level of 10, a ‘low-medium’ pain level of 30, a ‘medium’ pain level of 50, a ‘medium-high’ pain level of 70, and a ‘high’ pain level of 90. Level 100 was considered the point where the participant did not wish to experience a higher stimulation level in the experimental session and was not used. We repeated this procedure three times and computed the average stimulus intensities over these three repetitions corresponding to NPSs 10, 30, 50, 70, and 90.

Participants then underwent a pre-experiment test procedure: stimulus intensities corresponding to their pain intensity ratings NPS 10 to 90 were delivered in a pseudo-randomized order four times and participants were instructed to identify the intensity of each pulse. Participants had to correctly identify 75% of stimulus intensities to continue to the main experiment. If they did not achieve this in the test procedure, the intensities were adjusted, and the test was repeated until participants correctly identified 75% of stimulus intensities (Hird et al., 2019). Participants were excluded if their level of correlation between their ratings of the five levels of the nociceptive stimulus intensities within the baseline phase was below r=0.75, suggesting insensitivity to the varying nociceptive degrees of the electrical stimulus, and an essentially flat response profile. No participants were excluded with this criterion.

Having established the individual levels of intensity for each stimulation, participants were informed that their task was to rate the intensity and unpleasantness of nociceptive stimulations. We carefully described the distinction between intensity and unpleasantness ratings using the standard language developed by Price et al. (1989), emphasizing that pain intensity and unpleasantness should be rated independently.

The experiment started with a no-feedback session in which participants’ heartrate in anticipation of the shock (measured with ECG) and perceived level of pain (measured with Numeric Pain Scale and Likert scale) were assessed. These measures were used to normalize our dependant variables for our analysis (see Data analysis), and to create the acoustic false feedback stimuli to be used as feedback in the experimental session.

Participants were seated in front of the computer, with the ECG electrodes attached, their left hand placed on the table with the electrodes of the cutaneous electrical stimulation fixed on the phalanx of the middle finger, and their right hand placed on the mouse, ready to start the task. The ECG was recorded for the entire session.

Each trial of the no-feedback phase started with the presentation of a fixation cross appearing on the screen. After 60 seconds, a single electrical shock was administered, randomly varying the level of intensity in each trial – as established in the previous calibration phase (i.e., NPS 10, 30, 50, 70, and 90). After each stimulation, participants first rated the intensity of the painful stimulus (instruction: “How intense was the painful stimulation?”) on a continuous Numerical Pain Scale (NPS) appearing on the screen, from 0 (“not at all painful”) to 100 (“extremely painful”). They then rated the unpleasantness of the painful stimulus (instruction: “How unpleasant was the painful stimulation?”) on a 5-point Likert scale (1=no pain, 2=weak, 3=moderate, 4=severe, 5=extremely severe). The two scales were both presented on the screen in succession, and participants expressed their perceptual judgments through a mouse click on the scales. Between trials, a pause screen was shown for 45 seconds, in order to avoid the subsequent trial to be contaminated by the heart rate response to the shock. Then, a new screen appeared with the instructions to press the spacebar to start the new trial.

The no-feedback phase comprised 10 trials, after which participants were asked to wait and rest for 15 minutes before participating in the second experimental phase. This time was used by the experimenter to prepare the acoustic stimuli for use in the second session for each participant. For this purpose, we assessed the individual mean heart rate over the anticipation of the painful stimulus by measuring the mean R-R intervals of the ECG for the whole 1-min interval preceding the shock, and by transforming them into frequency (1/R-R, Colloca et al., 2006). The congruent feedback consisted of tones of heartbeats reproduced at the same individual HR frequency recorded over the no feedback phase. The incongruent faster and slower feedback consisted of tones of heartbeats reproduced at a frequency corresponding to an R-R interval obtained by either reducing or increasing the length of the original mean R-R interval over the no-feedback phase of 25%, respectively. Specifically, the incongruent acoustic feedback always started at a frequency rate that reproduced the individual heartrate recorded over the no feedback phase, to then gradually increase or decrease the R-R length according to whether the feedback was slower or faster, respectively.

In the feedback phase participants were informed that their task was similar to the previous no-feedback procedure, namely, to rate the intensity and unpleasantness of pain stimuli. Crucially, they were also informed that over the 60 seconds preceding the administration of the shock, they were exposed to acoustic feedback, which was equivalent to their ongoing heart rate. This phase consisted of 18 trials of 60 seconds (i.e., fixation) each, after which the electrical shock was administered, randomly varying its intensity in each trial (i.e., NPS 30, 50, and 70). Each level of intensity of the nociceptive stimulus was associated with one of the three different types of acoustic feedback, one congruent and two incongruent (i.e., slower and faster) with the participant’s heart rate recorded over the anticipation of the shock in the previous no-feedback phase. As in the previous no-feedback phase, participants’ individual mean heart rate in anticipation of the painful stimulus (i.e., ECG) and the perceived intensity of pain (i.e., NPS and Likert scale) were assessed. The order of the experimental conditions (i.e., congruent, slower and faster) was randomly generated for each individual participant by a web-based computer program (www.randomization.com)

Body Perception Questionnaire

After the experiment, all participants completed the Body Perception Questionnaire (Short form BPQ-SF) (Poli et al., 2021), to investigate whether either perceptual or autonomic modulation as induced by our task would be predictive of self-reported measures of bodily awareness and reactivity.

The Body Perception Questionnaire is a 22-item self-administered questionnaire that assesses awareness and reactivity of the autonomic nervous system, that is, the subjective ability to perceive bodily states and bodily reactions to stress. High scores on the BPQ reflect high awareness of internalbodily signals (i.e., high interoceptive sensibility) and high perceived reactivity of the visceral nervous system. Items ask participants to rate, on a 5-point scale (from 1 = never to 5 = always), the frequency with which they feel aware of bodily sensations (e.g., body awareness subscale “My mouthbeing dry”), experience supradiaphragmatic reactivity (e.g., supradiaphragmatic reactivity subscale “I feel shortness of breath”), and subdiaphragmatic reactivity (e.g., subdiaphragmatic reactivity subscale reactivity subscale “I have indigestion”). In this work, we focus on the body awareness and supradiaphragmatic reactivity subscales, given our specific focus on attention to the heart (Petzschner et al., 2019).

Data Analysis

Pain ratings (i.e., Likert and NPS ratings) collected in the feedback phase were normalized to the no-feedback phase in order to control for inter-individual variability in pain perception, as established by previous research (Bartolo et al., 2013; Cecchini et al., 2020; Riello et al., 2019) following the formula: Normalized value = (X-bX)/bX, where X represents the mean value of each measure assessed in the experimental feedback phase and bX the mean value of the measure calculated over the no-feedback phase. The same procedure of normalization was coherently applied to measures of the heart rate, recorded with the ECG in both the experimental feedback and no-feedback phase. We did not have specific hypotheses about how the different levels of noxious stimulus intensity affect the perceptual illusion, as we used them only to generate variability, thus all dependent variables were averaged across the three intensities (i.e., NPS 30, 50, and 70).

To provide a directional measure of whether the exposure to the faster and slower cardiac feedback induced similar or contrasting changes in actual heart rate and pain ratings, the mean value of all our dependant variables obtained over the exposure to the congruent feedback phase was subtracted to the scores acquired in both the incongruent faster and incongruent slower phase. This enabled us to obtain a variation rate (i.e., delta, Δ) for all our variables of interest (Δ pain intensity ratings, Δ pain unpleasantness rarings, Δ heart rate). Analysis was conducted on these values, allowing us to show whether the exposure to the false feedback manipulation (i.e., incongruent faster and slower feedback), relative to the exposure to the veridical (i.e., congruent) cardiac feedback, elicited a decrease or increase either in the perception of pain or in the actual cardiac frequency recorded over the anticipation of the nociceptive stimulus. No change in the incongruent feedback phase relative to the congruent feedback would produce a value of 0. Positive and negative values denote an increase or a decrease, relative to the exposure to the congruent feedback, either in pain ratings or anticipatory cardiac frequency, respectively. All data (i.e., Likert, NPS ratings and HR) were normally distributed (all p > 0.05) and analyzed with planned paired sample t-test, separately for each dependant variable, that is Δ Numeric Pain Scale, Δ Likert ratings and Δ heart rate parameters. The alpha criterion of significance of all tests was assumed at p < 0.05.

Results

Numeric Pain Scale of Intensity

Figure 2a shows the mean of Δ pain intensity ratings (Numeric Pain Scale ratings) after hearing the incongruent faster and slower exteroceptive feedback, relative to congruent feedback baseline. Participants’ mean values of Δ pain intensity ratings after faster vs. slower cardiac feedback were compared with a paired sample t-test. The results revealed that participants’ perception of pain intensity increased after the exposure to the faster (mean 0.10, SD 0.23) vs. slower (mean -0.01, SD 0.16) cardiac feedback, t(33) = 3.33, p = .002, d = .59 (Figure 2a). When compared to congruent feedback with additional simple t-tests, significant differences emerged for faster feedback, t(33) = 2.66, p = .01, d = .377, but not for the slower feedback, which showed a slight numerical decrease, t(33) = 0.47, p = .63, d = .034 (not shown).

Figure 1.

Figure 2.

Pain Unpleasantness Ratings Scale

The Δ pain unpleasantness ratings after the exposure to the slower and faster cardiac feedback, are shown in Figure 2b. Mean Δ pain unpleasantness ratings were analysed analogously. A paired sample t-test showed that hearing a faster vs. slower cardiac frequency affected pain unpleasantness ratings, t(33) = 2.77, p = .009, d = .51, such that Δ pain unpleasantness ratings were higher after faster feedback (mean 0.083, SD 0.15) than slower cardiac feedback (mean 0.005, SD 0.14) (Figure 2b). When compared to congruent feedback with simple t-tests, significant differences emerged for faster feedback, t(33) = 3.045, p = .004, d = .512, but not for the slower feedback, which showed a numerical slight decrease t(33) = 0.206, p = .83, d = 0.064 (not shown).

Real Heartrate

The Δ heart rate acquired over the exposure to the slower and faster cardiac feedback, relative to thecongruent feedback, is shown in Figure 2c. The paired sample t-test showed that, relative to the congruent cardiac feedback, participants’ real heart rate significantly decreased when hearing the faster (mean -0.007, SD 0.020) compared to the slower feedback (mean -0.02, SD 0.011), t(33) = 2.07, p = .045, d = .36 (Figure 2c). When compared with congruent feedback, significant heart rate differences emerged for faster feedback, t(33) = 2.28, p = .032, d = 0.21, but not for the slower feedback, which showed a slight numerical decrease t(33) = 0.92, p = 0.36, d = .05 (not shown).

Body Perception Questionnaire

The Short form Body Perception Questionnaire (Poli et al., 2021) administered at the end of the experiment allowed us to explore whether individual differences in body awareness and supradiaphragmatic reactivity were related to the extent to which participants developed (1) the perceptual illusion of pain (2) the physiological changes in their cardiac frequency. Overall, the average score obtained at the Body awareness subscale and Supradiaphragmatic subscale was 5.22 (SD= 1.62) and 2.05 (SD= 1.96), respectively. Each participant’s score was correlated with their level of Δ pain intensity, Δ pain unpleasantness ratings and Δ heart rate obtained in the faster feedback phase, which was key driver of effects, as seen also in prior research (Valin, 1996; Iodice et al 2019).

Pearson’s correlation showed no association between Δ heart rates recorded over the faster feedback phase relative to the congruent feedback and individual’s Body awareness subscale and Supradiaphragmatic subscale scores (all p > 0.05). Correlating individual’s Body awareness subscale and Supradiaphragmatic subscale scores with individual’s Δ pain unpleasantness did not show any association between the variables (all p > 0.05). Individuals’ Δ pain intensity ratings did not show any association with their Body awareness subscale scores (p > 0.05), but a negative correlation with the Supradiaphragmatic Reactivity (SUP) scale score (r = -.36, p = .033), such that higher Δ pain intensity ratings were associated with lower BPQ scores. This relationship however does not survive Bonferroni correction (p > 0.025).

Experiment 2

Experiment 1 established that false accelerated cardiac feedback induced both an illusory increase in the perception of pain and a modulation in the pain-anticipatory related cardiac state (i.e., heart rate) towards the expected level of the noxious stimulus, based on the frequency rate of the cardiac feedback. Experiment 2 tests whether the same perceptual and physiological changes are observed if participants are exposed to an accelerated non-interoceptive feedback (i.e., unrelated to biological human sounds), which should be less likely to induce expectations of threat. To address this question, we substituted the interoceptive cardiac feedback with an exteroceptive tone but kept the experiment otherwise identical. Thus, as in Experiment 1, participants’ pain intensity (i.e., Numeric Pain Scale), unpleasantness (i.e., Likert scale) ratings and cardiac frequency (i.e., ECG) were collected both in a no-feedback as well as in a feedback phase, which could be either congruent, faster or slower than participants’ individual heartrate recorded over the no-feedback phase. We reasoned that participants would have no expectation that an external acoustic sound signals a preparation of their body to a threatening, high painful stimulus. Consequently, perceptual and cardiac modulations associated to the feedback manipulation should be reduced over the exposure to the faster exteroceptive sound.

Methods

Participants

A sensitivity analysis with G*Power 3.1 (Faul et al., 2007) showed that a sample size of 29 provides .90 power to detect effects with Cohen’s d =.62. (SESOI of δ = 0.37). Thirty-five participants (mean age 25.17, SD= 4.02, 23 women) were recruited from Gabriele D’Annunzio University and the wider community. All were right-handed with normal or corrected-to-normal vision. As in Experiment 1, exclusion criteria for taking part were self-reported chronic and acute pain, neurological disease, serious cardiovascular disease (i.e., any type of disease involving the heart or blood vessels that might result in life-threatening medical emergencies, e.g., arrhythmias, infarct, stroke), or conditions that could potentially interfere with pain sensitivity (e.g., drug intake or skin diseases). Participants were asked to either not drink coffee or smoke cigarettes in the 60 minutes preceding the experiment. All gave written informed consent,were unaware of the purposes of the study, and were fully debriefed about it at the end of the experiment. Ethical approval from the local ethics board was obtained.

As in Experiment 1, participants were excluded if their correlation between the desired five levels of the nociceptive stimulus intensity and their ratings over the no-feedback condition was below r=0.75, suggesting insensitivity to the varying nociceptive degrees of the electrical stimulus. Five participants were excluded with this criterion. Moreover, participants were excluded if, for any reason, they linked the exteroceptive sound to a simulation of an interoceptive, cardiac acoustic feedback, in their post-experiment debriefing. One additional participant was excluded with this criterion. Hence, the final sample size was 29.

Apparatus and Stimuli

The apparatus and stimuli were the same as Experiment 1. The only difference regarded the selectionof the tone used for the creation of the feedback. Instead of using a single heartbeat, the sound was asingle percussion obtained by knocking two woods, gathered from https://freesound.org and manipulated in Audacity. The feedback audio was then created in MATLAB (R2020a), repeating the single tone according to the desired frequency (see Procedure).

Procedure

The procedure used in Experiment 2 was the same as Experiment 1, with the only exception that thesound used for the feedback was an exteroceptive, instead of interoceptive (i.e., cardiac) tone.

Data Analysis

Data analysis was identical to Experiment 1. Additionally, equivalence tests (i.e., TOST ‘two one-sided t-tests’ (TOST) procedure, (Lakens, 2017; Lakens et al., 2018) were performed for each variable (i.e., Δ pain unpleasantness and intensity ratings, Δ heart rate). Equivalence tests investigate whether the hypothesis that there are effects extreme enough to be considered meaningful can be rejected. We performed equivalence tests to examine whether the effect of feedback on pain-anticipatory heart rate and pain ratings was at least as extreme as the smallest effect size of interest (SESOI). If differences in our variables of interest at least as extreme as the equivalence bounds (i.e. here in standardized effect sizes) can be rejected in two-one sided tests (also known as one-tailed tests; i.e., the TOST procedure), we concluded that either pain ratings or pain-anticipatory heart rate in relation to the exposure to faster vs. slower feedback ratings are statistically equivalent, meaning the difference between our condition of interest (i.e. faster vs. slower feedback) is smaller than what is considered meaningful and statistically falls within the interval indicated by the equivalence bounds.

Results

Numeric Pain Scale of Intensity

Figure 3a shows the mean of Δ pain intensity ratings given after hearing the incongruent faster vs. slower exteroceptive feedback, relative to the congruent feedback baseline. Data were analysed as in Experiment 1, using a paired t-test to evaluate whether pain ratings differed after hearing the incongruent faster and slower feedback. In contrast to Experiment 1, mean Δ pain intensity ratings were not affected by feedback frequency, t(28) = .65, p = .51, d = .12 (Figure 3a). TOST procedure (Lakens, 2017) indicated that the observed effect size (d = .12) was significantly within the equivalence bounds of ΔL = –.57 and ΔU = .57, p = .011. Relative to congruent feedback, intensity ratings were increased for both faster and slower feedback, even though the difference was robust only for the faster feedback, t(28) = 2.18, p = 0.037, d = 0.28, not the slower cardiac feedback, t(28) = 1.58, p = 0.12, d = 0.22.

Figure 3.

Pain Unpleasantness Ratings Scale

Figure 3b shows the mean Δ pain unpleasantness ratings given after the exposure to the incongruent faster vs. slower cardiac feedback, relative to the congruent feedback. A paired sample t-test revealed no significant difference between Δ pain unpleasantness ratings after slower vs. faster feedback, t(28) = 1.17, p = .24, d = .19 (Figure 3b). TOST procedure (Lakens, 2017) indicated that the observed effect size (d = .19) was significantly within the equivalence bounds of ΔL = –.57 and ΔU = .57, p = .035. Further simple t-tests showed that, relative to the congruent feedback, differences did neither emerge for faster feedback, t(28) = 1.51, p = .14, d = .28, nor for the slower feedback, t(28) = .38, p = .70, d = .06 (not shown).

Real Heartrate

Figure 3c shows the Δ Heartrate acquired over the exposure to the slower and faster exteroceptivefeedback, relative to the congruent feedback. The paired-sample t-test revealed no significant difference between the exposure to the slower vs. faster exteroceptive feedback, t(28) = 1.48, p = .14, d = .26 (Figure 3c). TOST procedure (Lakens, 2017) indicated that the observed effect size (d = .26) was significantly within the equivalence bounds of ΔL = –.57 and ΔU = .57, p = .065. When heart rate recorded over the exposure to the slower and faster exteroceptive feedback was compared to the cardiac frequency acquired over the congruent feedback, no significant differences emerged for faster feedback, which showed a slight numerical decrease t(28) = .71, p = .48, d = .06, but for the slower feedback, which showed a robust decrease t(28) = -2.189, p = .037, d = .19 (not shown).

Additional comparison analysis between experiments

To address whether the perceptual and cardiac biases were differently modulated by the exteroceptive vs. interoceptive sound reproduced by the feedback, we directly compared the changes in either the anticipatory cardiac response or the reported levels of pain intensity and unpleasantness elicited by both feedback manipulations. We predicted that (1) the increase in both pain unpleasantness and intensity ratings observed after hearing the incongruent faster vs. slower feedback, and (2) the decrease in the real cardiac frequency observed over the exposure to the incongruent faster vs. slower feedback, were enhanced in the group of participants exposed to the interoceptive cardiac sound (Experiment 1), relative to the one that received the exteroceptive tone (Experiment 2). Such findings would demonstrate that perceptual and cardiac changes linked to the exposure to the accelerated feedback was largely associated to the generation of specific interoceptive expectations conveyed by simulated cardiac streams, which provided (false) information about the state of the body in relation to a threat.

To test this hypothesis, the Δ of pain ratings and heart rate parameters obtained after the exposure to the interoceptive (Experiment 1) and exteroceptive (Experiment 2) feedback, were entered into a 2 X 2 mixed measures analysis of variance (ANOVA) for the Δ Numeric Pain Scale ratings, Δ Likert ratings and Δ heart rate separately, with Experiment (Cardiac feedback vs. Exteroceptive feedback) as a between-subjects factor, and Feedback frequency (Faster vs. Slower) as within-subjects factors. Our prediction is that either the increase in the perceived pain intensity (Numeric Pain Scale) and unpleasantness (Likert Scale) or the heart rates decrease observed over the exposure to the faster vs. slower feedback will be reduced when participants were exposed to the cardiac feedback (i.e., Experiment 1), relative to the exteroceptive sound (i.e., Experiment 2). Interactions with Experiment will therefore reveal whether groups differ in how much either perceptual or cardiac frequency changes are biased by the type of sound reproduced by the manipulated feedback, and not simply the frequency rate of the acoustic feedback.

Numeric Pain Scale

The 2×2 ANOVA revealed a main effect of Feedback frequency F(1,61) = 8.28, p = 0.006, ηp² = 0.12. Participants’ reported levels of pain intensity were generally higher after exposure to the faster, relative to the slower feedback. Importantly, and as predicted, this main effect was qualified by an interaction of Feedback frequency and Experiment, F(1,61) = 4.04, p = 0.04, ηp² = 0.062. The faster feedback affected the perceptual bias more strongly in the Experiment 1 than in Experiment 2. Specifically, when participants were exposed to the faster, but not slower, cardiac interoceptive sound, their perceived and reported level of pain intensity was generally higher. This effect was not induced by sound reproducing an exteroceptive tone.

Pain Unpleasantness Ratings Scale

Results revealed a main effect of Feedback frequency F(1,61) = 8.11, p = 0.006, ηp² = 0.117, showing that after the exposure to the faster, relative to the slower feedback, participants generally reported higher levels of perceived unpleasantness associated to the noxious stimulus. This main effect was not qualified by interactions with Experiment (p = 0.13), showing that the specific type of sound reproduced by the acoustic feedback (cardiac vs. exteroceptive) did not differently modulated the perceived levels of pain unpleasantness associated to the nociceptive stimulus.

Heart rate

The 2×2 ANOVA showed no main effect of Feedback frequency F < 1. Importantly, the interaction Feedback by Experiment was significant F (1,61) = 6.253, p = 0.015, ηp2 = 0.093, showing that the exposure to the faster vs. slower cardiac feedback generally decreases heartrates, while listening to the exteroceptive faster vs. slower exteroceptive feedback generally increases participants’ cardiac frequency.

General Discussion

Interoceptive inference and the Embodied Predictive Coding frameworks (Barrett, 2017; Barrett & Simmons, 2015; Pezzulo, 2014; A. K. Seth, 2013; A. K. Seth & Friston, 2016) understand interoception as a reflection of the brain’s probabilistic inferences, through which it estimates and controls crucial physiological and homeostatic variables, by combining incoming sensory information from multiple modalities. In line with such views, expectations about pain should be integrated, in a Bayesian manner, with other interoceptive sources, here represented by cardiac streams reproduced with the acoustic feedback that generated predictions of increased threat levels associated with the forthcoming stimulus. Unbeknownst to participants, before the administration of the same intensity noxious stimuli, we varied the frequency of the cardiac feedback, manipulating the speed of what they thought was their ongoing heart rate. This sufficed to induce misleading expectations about the upcoming level of nociception associated with the stimulation.

In line with predictive models, reducing the gap between expected and actual interoceptive information can be achieved either by changing one’s perceptual state to encompass the new information, or by actively changing the body in response to the unexpected input, by engaging for example autonomous reflexes (i.e., perceptual and active inference, Parr et al., 2022, Friston, 2005; 2010; for a full discussion on interoception, see Ainley et al., 2016, Paulus et al., 2019). If so, then sensing an accelerated heartrate when anticipating pain might have two consequences. First, it may falsely lead participants to expect, and then perceive, a more intense pain stimulus then is really delivered. Second, it may induce changes in the internal states themselves, compensating for the perceived changes in heartrate by actually up-down-regulating one’s heartrate, towards the expected decreasing orienting cardiac response when expecting threatening stimuli (Lykken et al., 1972; Taggart et al., 1976; Bradley et al., 2005; Bradley et al., 2008; Colloca et al., 2012; Tracy et al., 2017). This inferential process has been demonstrated in various computational models of cardiac (Smith et al., 2020) and gastrointestinal (Smith et al., 2021) interoception, thermoregulation (Tschantz et al., 2022), and psychopathological conditions, such as anorexia nervosa (Barca & Pezzulo, 2020), panic disorder (Maisto et al., 2021), depression (Barrett et al., 2016; Stephan et al., 2016), and others (Paulus et al., 2019).

The results of Experiment 1 provide evidence for both components of the hypothesized response. They show, first, that faster cardiac feedback induces an illusory perception of increased unpleasantness and intensity of subsequent pain stimuli. Thus, as predicted, the mismatch between people’s expectations and the actual input (i.e., expecting a high level of pain but having equal nociceptive intensities of the stimulus) led to changes in the perception of pain, which reflected the expected input, more than the real changes that should be evoked by identical noxious stimulus. Secondly, the results showed that faster cardiac feedback also changed participants’ real heart rate. When hearing the faster feedback, the real heartrate tended to decrease, assuming the pattern of anticipatory response typically enacted when humans prepare to face a threatening stimulus (Lykken et al., 1972; Taggart et al., 1976; Bradley et al., 2005; Bradley et al., 2008; Colloca et al., 2006; Tracy et al., 2017). This evidence supports the hypothesis that not only interoceptive states modulate inferences, and then interoceptive perception, but the converse is also true: inference dynamics can modify or produce new internal states, to the extent that the actual cardiac frequency rate decreased as it would have done over the anticipation of a real threat.

Decreased cardiac activity in response to threat has been shown in both human and non-human species (for a review, see Livermore et al., 2021; Roelofs, 2017; Vila et al., 2007). More broadly, cardiac deceleration it is observed during highly salient events and attentive preparation (Jennings et al., 2009; Jennings & van der Molen, 2005; Van der Veen et al., 2000) and it is considered an evolutionarily-conserved, adaptive response that allows the organism to orient attention and enhance information processing and action preparation (Hashemi et al., 2019; Klaassen et al., 2021; Lojowska et al., 2015). Blending well with early cardiac reflexes literature, under which cortical excitation/inhibition is modulated by down- or up-regulating cardiac activity in order to engage or disengage, respectively, with salient stimuli (Lacey, 1970), recent proposals frames cardiac deceleration under a Bayesian inference approach. Under these views, the cardiac deceleration mechanism allows to adjust precision of sensory evidence accumulation relative to precision of bodily information, optimising perception and action (Skora et al., 2022). Thus, decelerated cardiac activity observed over the exposure to the faster feedback may reflect an autonomic adjustment to prioritize incoming noxious input information in light of the expectation of the threatening painful event.

To the best of our knowledge, this is the first study to successfully demonstrate both active and perceptual inference consequences of false interoceptive feedback, supporting the notion that, based on prior knowledge and expectations, the brain actively implements the interoceptive consequences that have been predicted (for a full discussion, see Barrett and Simmons, 2015; Pezzulo, 2014; Seth and Friston, 2016). Along similar lines, it has been demonstrated that is possible to generate predictions along the body dimension, inducing both active and perceptual inference. For example, in one of the most studied body illusion, the “rubber hand illusion”, the discrepancy between the seen and felt touch is reduced either by changing perception, such that individuals come to experience a sense of ownership over the rubber hand and misperceive the actual location of their hand (Botvinick & Cohen, 1998), or by active processes, such as performing automatic compensative movements (e.g., Asai, 2015) or engaging autonomic reflexes (e.g, Hohwy and Paton, 2010; Moseley et al., 2008). Interestingly, both perceptual and active inference during the rubber hand illusion can be induced also by expectation rather than experience of a tactile event (Ferri et al., 2013).

Our findings greatly expand upon this existing research, suggesting that, as it happens for bodily signals, the brain may analogously generate top-down interoceptive predictions with either perceptual or active consequences. Building upon emerging models that highlight the significance of active inference in addressing multisensory conflicts or accomplishing goals (Maselli et al., 2022), these findings have the potential to broaden the scope of these perspectives, encompassing domains ranging from motor control to interoceptive processes. In this regard, interoceptive and exteroceptive aspects of the body may draw upon similar predictive mechanisms, which involve the integration of diverse streams of information from multiple senses to form a central representation of their essential variables. This is in line with the emerging idea that the brain maintains an interoceptive schema for (optimal) multimodal integration. Paralleling the body schema (Head & Holmes, 1911), the interoceptive schema consists in a central representation of interoceptive variables (e.g., body temperature, cardiac activity etc.) along with prior assumptions or “set points” for these variables. This internal model would support homeostatic and allostatic regulation by optimally weighting multiple (i.e., proprioceptive, interoceptive, exteroceptive) streams of information to predict incoming interoceptive signals (Barca & Pezzulo, 2020; Iodice et al., 2019; Schoeller et al., 2022; Tschantz et al., 2022). If such a schema regulates our interoceptive states, the manipulation of illusions in this domain would serve as a unique framework to test it.

Not surprisingly, we did not find any counterpart of either the illusion of pain, or of the associated autonomous response, when providing participants with slower cardiac feedback. This result is in line with previous findings (e.g., (Iodice et al., 2019) that highlighted a similar asymmetry in the illusory perception of effort after a cardiac feedback manipulation. As in the current study, participants did not underestimate effort when provided with slower feedback, but they consistently overestimated effort levels after receiving faster cardiac feedback. Following this prior study, we interpret this result as a protective strategy through which predictive processes aim at promoting adaptive outcomes. Underestimating, but not overestimating pain constitutes a risk for the organism, as it can turn into a maladaptive behaviour and lead to dangerous consequences.

Importantly, both perceptual and cardiac frequency changes after faster (vs. slower) feedback were dramatically reduced when the heartbeat tone was replaced by an exteroceptive sound in Experiment 2. This finding supports the hypothesis that such perceptual and physiological modulations were specifically tied to the interoceptive streams reproduced by the feedback, which provided evidence of the current state of the body in preparation for a potential upcoming (threatening) stimulus. Therefore, varying the frequency of an external sound that is not informative of the current state of the body or related to interoceptive streams, should not induce any expectation about the nociceptive level of the stimulus. In other words, while hearing their own cardiac frequency varying may have been interpreted as a signal of their body preparation to a threatening painful event, participants would have no expectation about exteroceptive tone stimuli. Consequently, no discrepancy in the pain unpleasantness and intensity ratings, as well as in participants’ actual heart rates, were observed in Experiment 2, when participants were exposed to the faster, relative to the incongruent slower feedback, and the modulations induced in pain intensity and heart rates were significantly reduced.

Importantly, perceptual and autonomic changes were induced by false feedback about participants’ internal state, in contrast to prior work that manipulated expectations mainly with exteroceptive cues (Atlas et al., 2010; Colloca et al., 2006; Jepma et al., 2018; Tracy et al., 2017; Wiech et al., 2014), in line with the proposal that interoceptive inferences might integrate prior knowledge and all available sensory, proprioceptive and interoceptive signals that describe the state of the body (Iodice et al., 2019; Pezzulo, 2014). In particular, the findings provided here support the notion that pain represents a critical variable within the predictive coding scheme, and that its estimation exploits prior knowledge and beliefs, and, among others, devotes special attention to cardiac streams from within the body, which can lead us to conclude that the internal generative model of pain considers heart rate in the likelihood model. In turn, the estimate of pain and of other physiologically relevant variables might be key for allostatic functions and adaptive, energy-efficient regulation by the brain (Barrett et al., 2016).

Furthering the underlying mechanisms of illusions provide a precious opportunity to learn about the role of predictions in shaping our understanding of the world around and inside us, offering the occasion to observe it through the lenses of the internal generative models that guide perceptions and actions. This evidence provides a point of departure to develop a full picture of predictive interoceptive dynamics, as which are the interoceptive channels that the brain monitors to estimate crucial variables, or how predictive processes orchestrate all different streams of information within the embodied multisensory sphere. In this view, it is possible to create novel illusions and manipulate the precision of signals coming from different channels (i.e., proprioceptive, interoceptive, exteroceptive) to study whether the brain might weight and integrate specific sources of information to predict precise homeostatic variables.

A central assumption in predictive models is that predictions and associated prediction-error are represented in terms of precision, wherein ‘precision’ refers to the salience and reliability of each probability distribution (Ainley et al., 2016; Feldman and Friston, 2010; Hohwy, 2012; Palmer et al., 2019; Parr et al., 2022; Seth and Friston, 2016; Yon and Frith, 2021). In this regard, future studies may investigate whether the reliability of people’s expectations about the link between their heart rate and threat anticipation underpins the development of the interoceptive illusion. This might involve self-report measures that properly investigate and quantify participants’ beliefs, and how they manifest in perceptual or bodily changes. Exploring the degree of certainty of priors may constitute a testing bed to investigate the role of precision in weighting the available information that arises from different modalities. Future research may explore whether the illusion in pain perception might be altered by changing people’s prior knowledge about the link between their cardiac frequency and the experience of pain (e.g., providing participants some prior knowledge about potential decreases in heart rate in anticipation of pain). This may shed light on whether perceptual biases in the experience of pain are associated to overlearned priors that do not update to incoming sensory data, or, instead, whether it is possible to directly act on the perceptual illusion by changing prior knowledge. This would have far-reaching implications, resulting in the striking consequence that is possible to modify the perception of our internal milieu, but also impact autonomic states by simply reconfiguring individuals’ beliefs and knowledge. Such findings would hold profound consequences for understanding and potentially intervening in psychopathological conditions influenced by distorted perception and maladaptive cognitive biases.

In this regard, the contribution of the study of illusions offers important insights not only for a converging model of brain functioning under the principle of optimal predictive architectures, but it also provides implications for understanding psychopathology in terms of aberrant predictions and compounding allostatic consequences (for a full discussion, see Barrett & Simmons, 2015), based on the inability to accurately predict and revise, when needed, prior – interoceptive - beliefs.

In this view, the illusion of pain highlighted here, as well as other interoceptive illusions, may constitute an opening window on the understanding of disrupted balancing between prior knowledge and incoming sensory information, potentially explaining psychopathological conditions.

Acknowledgements

EP, FF, MC, GP and PB designed the experiments. Experiments were coded by EP and MGP. Data was collected by EP. Data were analysed by EP. The manuscript was written by EP, FF, MC, MGP, GP and PB. All authors gave final approval for publication. The work was funded by Leverhulme Trust grant RPG-2019-248 to PB, and PhD studentship awarded to EP from the Universities of Plymouth and Aberdeen. This research received funding from the European Union’s Horizon 2020 Framework Programme for Research and Innovation under the Specific Grant Agreement 945539 (Human Brain Project SGA3) and 952215 (TAILOR) to GP, the European Research Council under the Grant Agreement No. 820213 (ThinkAhead) to GP and the PNRR MUR project PE0000013-FAIR to GP. This work was supported by the “Departments of Excellence 2018–2022” initiative of the Italian Ministry of Education, University and Research for the Department of Neuroscience, Imaging and Clinical Sciences (DNISC) of the University of Chieti-Pescara, and by the “Search for Excellence” initiative of the University of Chieti-Pescara