Abstract
Pain perception arises from the integration of prior expectations with sensory information. Although recent work has demonstrated that treatment expectancy effects (e.g., placebo hypoalgesia) can be explained by a Bayesian integration framework incorporating the precision level of expectations and sensory inputs, the key factor modulating this integration in stimulus expectancy-induced pain modulation remains unclear. In a stimulus expectancy paradigm combining emotion regulation in healthy male and female adults, we found that participants’ voluntary reduction in anticipatory anxiety and pleasantness monotonically reduced the magnitude of pain modulation by negative and positive expectations, respectively, indicating a role of emotion. For both types of expectations, Bayesian model comparisons confirmed that an integration model using the respective emotion of expectations and sensory inputs explained stimulus expectancy effects on pain better than using their respective precision. For negative expectations, the role of anxiety is further supported by our fMRI findings that (1) functional coupling within anxiety-processing brain regions (amygdala and anterior cingulate) reflected the integration of expectations with sensory inputs and (2) anxiety appeared to impair the updating of expectations via suppressed prediction error signals in the anterior cingulate, thus perpetuating negative expectancy effects. Regarding positive expectations, their integration with sensory inputs relied on the functional coupling within brain structures processing positive emotion and inhibiting threat responding (medial orbitofrontal cortex and hippocampus). In summary, different from treatment expectancy, pain modulation by stimulus expectancy emanates from emotion-modulated integration of beliefs with sensory evidence and inadequate belief updating.
Significance Statement
Although evidence indicates that human pain experiences stem from the integration of prior beliefs with sensory evidence, the key factor that biases this integration process as pain perception is affected by stimulus expectancy remains unclear. Using fMRI, we examine participants’ emotional reactions when they expect and then perceive incoming painful stimuli. We find that the elicited anxiety during pain enhancement by expectations of increased pain influences the integration of expectations with sensory inputs and appears to impair the updating of expectations. By contrast, this integration mainly involves pleasant emotion provoked during pain attenuation by expectations of decreased pain. This work highlights the important role of emotions when human experience of pain is shaped by stimulus expectancy.
Introduction
Expectations are powerful cognitive factors that modulate human pain experiences, with negative expectations (expectations of increased pain) and positive expectations (expectations of decreased pain) enhance and alleviate pain, respectively (Shih et al., 2019). While anticipating pain, negative expectations evoke anxiety (Simpson et al., 2001), whereas positive expectations reduce the aversiveness of noxious stimuli and elicit rewarding responses (Seymour et al., 2005; Ellingsen et al., 2013). When stimulus expectancy influences pain processing, previous studies have reported activations within brain structures implicated in emotion processing (Lindquist et al., 2016), such as the amygdala (Atlas et al., 2010; Ziv et al., 2010), medial prefrontal subregions including the medial orbitofrontal cortex (mOFC) and anterior cingulate cortex (ACC) (Koyama et al., 2005; Keltner et al., 2006; Atlas et al., 2010), and hippocampus (Ploghaus et al., 2001; Ziv et al., 2010). These observations suggest the possibility that pain modulation by stimulus expectancy encompasses an emotional component; that is, the emotional responses that occur during this modulatory process may shape individual pain experiences. However, the neural mechanism underlying the influence of emotion on pain modulation by stimulus expectancy remains unclear.
There are two possible accounts of how this emotional component, if present, modulates pain perception. In terms of the theory on expectancy and perception, a Bayesian integration framework – which assumes that top–bottom prior beliefs are integrated with bottom–top sensory evidence, and both are weighted by their relative precision to form human perception (Friston, 2010; de Lange et al., 2018) – has been demonstrated to explain pain modulation by treatment expectations (e.g., placebo hypoalgesia) (Grahl et al., 2018). Since anticipating imminent pain causes emotional responses described above, and both negative and positive emotional information receives priority in processing (Vuilleumier, 2005; Gupta, 2019), it is plausible that, relative to the Bayesian integration, an emotion-associated integration of expectations with sensory inputs may better account for pain modulation by stimulus expectancy. This notion is supported by previous observations that the perceived threat of imminent pain and an individual's pain-related emotions can affect the relative influence of expectations and sensory inputs on pain perception (Lim et al., 2020), and the abovementioned amygdala and mOFC are involved in comparing expectations with sensory inputs (Gottfried et al., 2003; Saddoris et al., 2005; Seymour et al., 2005; Iordanova et al., 2021).
The second account relates to learning in threatening contexts. Pain is a perceived threat to one's biological integrity, and optimal updating of pain predictions by leveraging prior experiences not only is essential for survival (Tovote et al., 2015; Strickland and McDannald, 2022) but affects future pain experiences (Buchel et al., 2014; Geuter et al., 2014). Concerning the link between learning and emotions, it has been reported that emotions can bias behavioral and neural responses related to prediction errors (Watanabe et al., 2013) and the mismatch between expected and actual outcomes to update predictions (Schultz et al., 1997; Sutton and Barto, 1998). Negatively valenced emotions, such as anxiety, have been shown to prevent optimal updating of aversive outcome predictions (Browning et al., 2015; Piray et al., 2019). Such observations suggest that the emotional component of stimulus expectancy effects on pain, as proposed above, may also modulate pain perception via affecting the updating of pain expectations.
In this work, we aimed to characterize the role of emotion in pain modulation by stimulus expectancy. To this end, we employed an emotion regulation (ER) strategy to downregulate participants’ emotional responses arising from anticipation of impending pain and then assessed corresponding changes in pain perception. To further investigate how emotions modulated pain processing, we tested an emotion-associated integration of expectations with sensory inputs, and then assessed the prediction error of pain and updating of pain expectations. We hypothesized that expectations and sensory inputs are weighted by their associated emotions to underlie pain perception, with inadequate updating of negative expectations for pain due to emotion-associated suppression of prediction error signals.
Materials and Methods
Participants
Thirty-six right-handed, pain-free participants with neither prior experiences in pain-related experiments nor self-reported neuropsychiatric disorders took part in this fMRI experiment. Data from five participants were excluded: one could not differentiate between the three intensities of painful stimulation during the conditioning session, and four did not report graded expected pain intensity across the three pain-predictive cues after the conditioning session. This yielded a final sample of 31 participants (16 females; aged 20–31 years, mean ± SD = 22.6 ± 3.0 years) included in our behavioral validation analyses (see Results) and behavioral and imaging analyses associated with positive expectations. For negative expectation, data from additional three participants were excluded due to no negative expectations [one participant, whose expected pain intensity for the high-pain cue before the start of the test session was lower than that for the Attend MM trials (i.e., a medium-pain stimulus preceded by a medium-pain cue) in the test session] or no negative expectancy effects on pain in Attend trials [two participants, whose mean pain rating in the test HM trials (i.e., a medium-pain stimulus preceded by a high-pain cue) was lower than that for test MM trials], leaving 28 participants for analysis. For the skin conductance response (SCR) data, two participants were excluded due to technical problems, leaving 29 participants for analysis. The determination of sample size in the present study was based on three considerations. First, to evaluate the necessary number of participants required for the effect of ER on stimulus expectancy effects, we conducted a behavioral pilot study (12 participants; 10 females; aged 20–28 years, mean ± SD = 23.8 ± 2.6 years), in which we used the same paradigm as in the formal experiment and detected an effect size of d = 0.67 or 0.69 for negative or positive expectancy effects, respectively. A power analysis using G*Power (Faul et al., 2007) demonstrated that a sample size of 26 would allow for the detection of this effect size with 90% power and an α of 0.05 (two-tailed). Second, the cue-based expectancy protocol of the present study was based on a similar study published at our lab (Shih et al., 2019), in which we analyzed data from 23 participants and detected the involvement of ACC and aIC in pain modulation by stimulus expectancy. As such, we thought that a sample size of n ≥ 26 would allow us to detect significant activation within expectancy-associated brain structures. Third, because the present study compared the emotion integration model with the precision integration model (i.e., Bayesian integration; see below for details), the sample size was also chosen to approximately match a previous study (n = 31) that investigated pain modulation by expectations using the precision integration model (Grahl et al., 2018). The ethics committee of the National Taiwan University Hospital approved the study, and all participants gave written consent.
Stimuli
Electrocutaneous stimuli were generated by a bipolar constant-current stimulator (DS5, Digitimer) and delivered to the dorsum of the participant's left hand via a pair of MRI-compatible leads (LEAD108, Biopac Systems) and silver chloride surface electrodes (EL-508, Biopac Systems). Each stimulus consisted of 200 ms monophasic rectangular pulses with pulse duration of 2 ms and interpulse interval of 48 ms (i.e., 20 Hz). Stimulus presentation and behavioral data acquisition were implemented using Presentation (Neurobehavioral Systems) and LabVIEW software (National Instruments).
Experimental design
Behavioral session
In the beginning of the experiment, each participant's detection threshold for just noticeable sensation was first determined using an ascending method of limit protocol (start: 0.1 mA; step: 0.1 mA; interstimulus interval: 20 s). Afterwards, we defined three intensities of electrical stimulation that elicited pain scores of 25 (low pain), 50 (medium pain), and 75 (high pain) on a 0–100 numerical rating scale (0, no pain; 100, unbearable pain) for each participant. For this purpose, we preliminarily estimated the three stimulus intensities by applying two additional series of ascending stimuli (start, detection threshold; step, 0.6 mA; interstimulus interval, 20 s) and asking the participant to rate perceived pain on a numerical rating scale (0, no pain; 100, unbearable pain) after each stimulation. To more precisely determine the three target intensities, we subsequently performed a calibration procedure, in which we applied each of the three preliminarily estimated intensities and adjusted the stimulation intensity based on their pain rating on the numerical rating scale. If the rating was higher (or lower) than the target pain score, the intensity would be reduced (or increased) by 0.4 mA until the rating became lower (or higher) than the target pain score. At that point, the stimulus would then be increased (or reduced) by 0.2 mA until the next turn point. The average of the last two intensities that produced a higher- and lower-than-target pain score was taken as the stimulus intensity for that target pain score in each participant. To avoid habituation or sensitization to pain during the experiment, this calibration procedure was repeated before the start of each test session.
Next, participants took part in a training session to familiarize themselves with the ER strategy. Given the pain-predictive cues in the present study served as conditioned stimuli, we adopted an imagery-focused regulation strategy in our experiment, which has been demonstrated to be able to successfully downregulate cue-elicited emotions (Delgado et al., 2008a,b). Before the start of training, participants were informed that each training trial consisted of the display of a regulation instruction (Attend or Regulate) followed by a color cue (a blue, green, or yellow circle) that would later be predictive of specific pain intensity in the test session (described below). In the Attend condition, participants were instructed to simply fixate on the color cue. In the Regulate condition, they were prompted to imagine a calming thing corresponding to the color of the cue (i.e., “think of a thing corresponding to the color of the cue that calms you down, such as a blue sky for a blue circle”). During the training session, participants completed ten training runs, with each run containing all the three cues presented in random order. At the end of these runs, the experimenter verbally confirmed participants’ ability to correctly perform this imagery-focused strategy. They were only qualified to proceed to the following conditioning session if they reported at least five things that were matched to the cue in the color dimension. To ensure that participants also correctly performed this regulation strategy during scanning, this qualification process was repeated when they completed the test session.
Conditioning session
Prior to the test session, we administered a conditioning session, in which participants learned the contingencies between three color cues and their individualized high-, medium-, and low-pain stimulus (i.e., HH, MM, and LL trials; Fig. 1A). The cue-to-stimulus assignment was counterbalanced across participants, and the entire session consisted of six repetitions of each trial type (i.e., totally 18 trials) presented in a random fashion. The trial structure was the same as that in the test session (Fig. 1B), except that there was no 1.5 s instruction period. Each trial began with the presentation of a pain-predictive cue (4 s), which was followed by a 2.5–4.5 s delay interval (jittered). Then, a painful stimulus was delivered and, after a 1.3 s poststimulus interval, participants rated perceived pain on a 0–100 visual analog scale (VAS; 0: no pain; 100: unbearable pain) within 3.5 s. The inter-trial interval was jittered between 5 and 7 s. Participants were instructed to establish the cue–stimulus relationships during this session, which also served to familiarize themselves with the experimental procedures in the following test session.
Experimental design. A, Overview of the experiment. Following a behavioral session, the conditioning session included only correctly signaled trials (e.g., a high-pain stimulus preceded by a high-pain cue, i.e., an HH trial). The subsequent test session consisted of two scanning runs and included both correctly and incorrectly signaled trials (e.g., a medium-pain stimulus preceded by a high-pain cue, i.e., an HM trial). The negative (or positive) expectancy effect was assessed by the comparison between the HM (or LM) and MM trials. For each pain-predictive cue, we asked participants to report expected pain intensity before (Pre-Test) and after (Post-Test1 and Post-Test2; for both Attend and Regulate conditions) scanning. At these three time points, they also reported the overall anxiety and pleasantness linked to each pain-predictive cue. H, high; L, low; M, medium. B, Trial structure in the test session (the conditioning session had the same trial structure except for no regulation instruction). Each trial was initiated with a regulation instruction (Attend or Regulate; 1.5 s) followed by a pain-predictive cue (high, medium, or low pain; 4 s). After an anticipation period (2.5–4.5 s), an electrical pain stimulus (0.2 s) was applied to the dorsum of the participant's left hand. Finally, after a poststimulus period (1.3 s), participants rated perceived pain using a visual analog scale within 3.5 s. The intertrial interval (ITI) was 5–7 s.
Test session
The test session was composed of five trial types [three correctly signaled trials (HH, MM, and LL) and two incorrectly signaled trials (HM and LM); Fig. 1A] in two regulation conditions (Attend and Regulate). The correctly signaled trials served to maintain participants’ belief about the cue-stimulus contingencies. Unbeknownst to the participants, the incorrectly signaled trials enabled us to examine stimulus expectancy effects on pain. The trial structure was identical to that in the conditioning session but with an additional regulation instruction period (1.5 s; “Attend” or “Regulate”) presented at the beginning of each trial (Fig. 1B). In the Attend condition, participants were instructed to think about the predicted pain intensity pertaining to each cue without regulating cue-elicited emotions (i.e., “think of the predicted pain intensity of this cue”). In the Regulate condition, they voluntarily reduced cue-elicited emotions using the same procedure as described above in the behavioral session. They were reminded to use the learned imagery ER strategy to the best of their ability, and were asked to rate the experienced pain intensity for each painful stimulus. For the isolation of neural substrates supporting expectancy effects on pain, the number of HM, MM, and LM trials were equal (each with six repetitions per regulation condition per session). In addition, there were three repetitions of HH and LL trials per regulation condition per session. Consequently, there were 48 trials presented in randomized order in each session, and the entire test session consisted of two scanning runs (Test1 and Test2; Fig. 1A). Note that, in order to consolidate participants’ belief about the cue-stimulus contingencies following the conditioning session, the first six trials in each test session consisted of the three correctly signaled trials in Attend and Regulate conditions.
To check whether participants correctly learned the cue-stimulus contingencies from the conditioning session, we asked them to report their mean expected pain intensity for each cue using a 0–100 VAS (0, no pain; 100, unbearable pain) before the start of the test session (Pre-Test; Fig. 1A). To examine whether ER influenced pain expectations during the test session, we also asked them to report cue-associated pain intensity in both Attend and Regulate conditions during scanning after each scanning session (Post-Test1 and Post-Test2; Fig. 1A). Before receiving painful stimulation in each trial, participants needed to both prepare for ER (following the Attend or Regulate instruction) and anticipate a specific pain intensity (Fig. 1B, following the color cue). To avoid interference effects on these cognitive processes, we therefore did not ask participants to report trial-by-trial expected pain intensity in this fMRI experiment.
For the assessment of emotion, a discrete emotion encompasses different measurable parts (such as subjective experience, physiological responses, expressive behavior, etc.), with different measures appearing sensitive to different aspects of the emotional state (Mauss and Robinson, 2009). In the present study, we measured each participant's subjective experience (self-reports of current emotional experiences) as well as autonomic changes (SCR) commonly used to track the arousal level of the emotional state (Lang et al., 1993; Critchley, 2002; Mauss and Robinson, 2009). Therefore, the word “emotion” in the present study, especially in the emotion integration model (described below), mainly refers to subjective self-report. To confirm whether ER was capable of reducing participants’ emotional responses during scanning, we (1) asked participants to report the overall anxiety and pleasantness levels for each cue based on cue-associated pain experiences using a 0–100 VAS for anxiety and another for pleasantness (0: no anxiety or pleasantness; 100: very anxious or pleasant) in both Attend and Regulate conditions before and after each scanning session (Pre-Test, Post-Test1 and Post-Test2; Fig. 1A) and (2) continuously recorded participants’ SCR during scanning (described below). In this fMRI experiment, we did not collect participants’ anxiety and pleasantness rating for each trial, because previous studies have reported that trial-by-trial emotion rating during scanning may reduce activation within emotion-related brain structures (Hariri et al., 2003; Lieberman et al., 2011; Dorfel et al., 2014).
SCR data collection and analysis
SCR was recorded during the test session using a Biopac MP150 data acquisition system (Biopac Systems, Inc) and AcqKnowledge software, with MRI-compatible electrodes (EL509; Biopac Systems) placed at the intermediate phalanges of the middle finger and ring finger of the participant's left hand. The sampling rate was 2 kHz, the gain set to 5 microSiemens (mS)/V, and the low pass filter set to 1.0 Hz. SCR data were down-sampled to 250 Hz and processed using the Autonomate software in MATLAB (MathWorks), which provides automated scoring of stimulus-locked SCR responses (Green et al., 2014). The SCR was considered valid if the trough-to-peak deflection started between 0.5 and 4.5 s following the cue onset, lasted between 0.5 and 5.0 s, and was greater than 0.02 μS. Trials that did not meet these criteria were scored as zero. To reduce interindividual variability to facilitate statistical analysis, SCRs were then square root-transformed and averaged for each cue in each regulation condition (Boucsein, 2012). To assess the effect of ER on stimulus expectancy-associated emotional responses during scanning, we subtracted the mean SCR amplitude elicited by the medium-pain cue from that elicited by the high- and low-pain cue between Attend and Regulate conditions.
Behavioral analysis and computational modeling
In the present study, individual magnitude of stimulus expectancy effect on pain was calculated by comparing the mean pain rating in MM trials with that in HM trials (“HM − MM” for negative expectations) or LM trials (“MM − LM” for positive expectations). The difference in this magnitude between Attend and Regulate conditions was defined as the effect of ER on stimulus expectancy effect.
In a Bayesian integration framework (Buchel et al., 2014), participants’ perceived pain (i.e., the pain ratings in test HM or LM trials; posterior) emanated from the integration of previous stimulus expectations (prior) with incoming sensory inputs (likelihood), with these two Bayesian components represented in terms of Gaussian probability density functions:
Hypothetical precision and emotion integration models and Bayesian model comparison. A,B, In the precision integration model, prior stimulus expectations [i.e., the mean pain ratings in conditioning HH trials for negative expectations (shown in blue; A) or conditioning LL trials for positive expectations (shown in yellow; B);
The mean posterior [i.e., the hypothetical pain perception for the test HM (or LM) trials] and the predicted magnitude of stimulus expectancy effect (i.e., the difference between the mean intensity of posterior and likelihood) can thus be estimated using the following equations:
As described above, the predicted magnitude of stimulus expectancy effect according to the emotion integration model can be estimated using Equations 5 and 6.
Finally, we investigated whether the precision or emotion integration model better explained the obtained data. We first calculated each participant's value for marginal likelihood in each model, the probability of the data given a particular model (Demichelis et al., 2006). The Bayes factor (BF) defined by the ratio of the marginal likelihood between these two models was then used to quantify the support for one model compared to another (i.e., BF12 for Model 1 over Model 2 and BF21 for Model 2 over Model 1), with a BF > 3 indicating at least moderate evidence in favor of the first given model (Kass and Raftery, 1995). Based on the assumption that each of the two models was equally likely a priori, each participant's BF was then transformed into the posterior model probability for each model [e.g., BF12/(BF12 + 1) for Model 1 over Model 2]. Subsequently, to examine the relative plausibility of these two models at the group level, we performed random effects Bayesian model selection [using the spm_BMS function in Statistical Parametric Mapping (SPM)] to estimate the expected posterior probability (i.e., expected frequency) for each model (i.e., the probability that a given model generates the gathered data of randomly selected subject) and exceedance probability (the possibility that one model is more likely than another in the comparison) (Stephan et al., 2009; Rigoux et al., 2014).
To verify the involvement of a prediction error mechanism in the effect of emotion on expectancy-associated pain modulation, we need to confirm a significant difference in pain prediction errors between the Regulate and Attend conditions. For this purpose, the difference between the mean pain rating (average of Test1 and Test2) and mean expected pain intensity (average of Post-Test1 and Post-Test2) during scanning was defined as the mean prediction error of pain for each trial type (Schenk et al., 2017; Shih et al., 2019), given expectancy effects have been shown to remain constant across the entire experiment in previous pain expectation studies (Schenk et al., 2017; Shih et al., 2019). Also, as described above in the “Experimental design” section, we did not ask participants to report trial-by-trial expected pain intensity, so as to avoid interference effects on their cognitive processes.
Statistical analysis
We used GraphPad Prism software to perform statistical tests. A paired t test was used to compare the emotion ratings and stimulus expectancy effects on pain between Attend and Regulate trials. For the stimulus expectancy effects on pain, SCRs, and prediction error of pain, a 2 (expectation: negative or positive) × 2 (instruction: Attend or Regulate) repeated-measures ANOVA was employed to examine the effect of ER. Another 2 (expectation: negative or positive) × 3 [condition: Pre-Test, Attend trials (average of Post-Test1 and Post-Test2), or Regulate trials (average of Post-Test1 and Post-Test2)] repeated-measures ANOVA was conducted to examine the effect of ER on expected pain. All repeated-measures ANOVAs were followed by Bonferroni post hoc tests correcting for multiple comparisons. Pearson's correlation was used to test for associations between two continuous variables. Unless otherwise specified, p values refer to two-tailed tests in the present study. Given our strong unidirectional predictions that (1) negative expectancy effects on pain may be positively predicted by participants’ anxiety levels, (2) positive expectancy effects on pain may be positively predicted by their pleasantness levels and/or negatively predicted by their anxiety levels, and (3) the hypothetical stimulus expectancy effect (Eqs. 5, 6) would positively correlate with the observed stimulus expectancy effect, we applied one-tailed tests in relevant correlation analyses.
Imaging data acquisition and analysis
Brain images were recorded at a 3-T Magnetom Prisma MR system (Siemens) using a 64-channel head coil. T2*-weighted blood oxygen level-dependent (BOLD) images were acquired using a gradient-echo echo planar imaging (EPI) sequence (37 contiguous axial slices; TR = 2,000 ms; TE = 30 ms; flip angle = 90°; FOV = 224 × 224 mm2; a GRAPPA acceleration factor of 2; slice thickness = 3.9 mm; voxel size = 3.5 × 3.5 × 3.9 mm3; acquisition matrix = 64 × 64). The initial four EPI volumes were discarded to allow for steady-state magnetization. For registration purposes, participants also underwent high-resolution structural T1-weighted magnetization-prepared rapid acquisition gradient echo (MP-RAGE) scans (voxel size = 0.88 × 0.88 × 0.89 mm3) and T2-weighted scans that were coplanar to the EPI but with higher in-plane resolution (256 × 256). A magnetic field map was acquired by using a double gradient-echo sequence (TR = 600 ms; TE1 = 10.00 ms; TE2 = 12.46 ms).
All imaging data were analyzed using SPM12 software (Wellcome Centre for Human Neuroimaging). The EPI volumes were unwarped using the field map to compensate for magnetic field inhomogeneities (Hutton et al., 2002), realigned to the first image of each session to correct for motion, and corrected for differences in slice acquisition timing. The output mean EPI image was then coregistered with the participant's T2-weighted structural image, which in turn was aligned with the T1-weighted image. Subsequently, the coregistered T1-weighted images were segmented to gray matter, white matter, and cerebrospinal fluid according to the International Consortium for Brain Mapping East Asian brain template to create a study-specific template using the Diffeomorphic Anatomical Registration Through Exponentiated Lie algebra (DARTEL) toolbox. This procedure also generates “flow fields,” which parameterize the nonlinear deformations in each participant and were used to normalize each participant's realigned and resliced EPI images to the standard Montreal Neurological Institute (MNI) space (voxel size: 2 × 2 × 2 mm3). Finally, images were spatially smoothed using a Gaussian kernel full-width at half-maximum of 6 mm. In this work, statistical parametric maps are overlaid on the T1-weighted images averaged across participants.
The normalized and smoothed functional images were analyzed using the general linear model (GLM) in SPM12. In each participant's first-level GLM, the main regressors of interest consisted of the 10 epoch regressors [five trial types (HH, HM, MM, LM, and LL) under two instruction types (Attend and Regulate)] representing BOLD activation during painful stimulation (duration: 0.2 s). Other regressors in this GLM included one regressor for the instruction (collapsed across all conditions; duration: 1.5 s), six regressors for the cue period (duration: 4 s; 3 cues × 2 instruction types), six regressors for the delay interval (duration: 2.5–4.5 s; 3 cues × 2 instruction types), 1 regressor for the rating period (collapsed across all conditions; duration: 3.5 s), and six regressors for head motion parameters. All regressors were convolved with a canonical hemodynamic response function (Friston et al., 1995), and the data were high-pass filtered with a frequency cutoff at 128 s to remove low-frequency drifts. Each participant's first-level t-contrasts of interest were subsequently used for second-level GLM analyses (Holmes and Friston, 1998).
To validate our approach, we first used whole-brain analyses to analyze the BOLD response for the contrasts (1) “Regulate > Attend” (pooled over the three pain-predictive cues) during the cue period to examine activation in brain regions implicated in ER and (2) “HH > LL” during painful stimulation in the conditioning session – which minimizes the potential bias by expectations and ER – to examine activations within brain regions reflecting pain intensity. Following our previous work, we also performed small-volume correction (SVC) analyses on the contrast “HM > MM” (or “LM > MM”; pooled over Attend and Regulate conditions) to test whether we could replicate the engagement of the ACC [or anterior insular cortex (aIC)] in pain modulation by negative (or positive) expectations reported in our previous study (Shih et al., 2019). Next, our fMRI data analysis consisted of three steps. First, to investigate whether the amygdala (for negative expectations) and mOFC and anterior hippocampus (aHPC) (for positive expectations) encoded the emotional component of pain modulation by stimulus expectancy, individual brain responses associated with pain modulation by negative (or positive) expectations [i.e., contrast “HM > MM” (or “LM > MM”)] during the stimulation period in Attend versus Regulate trials were regressed against corresponding anxiety (or pleasantness) ratings for the high- (or low-) pain cue [i.e., (AttendHIGH or LOW – AttendMEDIUM) – (RegulateHIGH or LOW – RegulateMEDIUM)]. In our participants, we found that ER-induced emotional changes were counterproductive (i.e., ER increased participants’ emotional ratings) in six participants for negative expectations and 11 participants for positive expectations. In these participants, ER did not significant influence stimulus expectancy effects on pain (both p ≥ 0.301). This indicates that an increase in anxiety (for negative expectations) or pleasantness (for positive expectations) after ER did not significantly influence negative or positive expectancy effects on pain, respectively. Since our aim was to search for brain regions whose activity not only reflected ER-induced reduction in cue-elicited emotions but modulated pain processing, it was thus unnecessary that activity in these regions should also reflect ER-induced increase in cue-elicited emotions. As such, this covariate was set to zero in these participants. To correct for the number of regions of interest (ROIs), the mOFC and aHPC ROIs were combined into a single mask for the SVC analysis in positive expectancy effects. Second, we examined whether the functional interaction between the amygdala and ACC underlay the integration of negative expectations and sensory inputs, and whether the mOFC interacted with the aIC and/or aHPC to underlie the integration of positive expectations and sensory inputs (see Psychophysiological interaction analysis section below). Third, given our fMRI data showed that coactivation between the left amygdala and ACC reflected anxiety-associated posterior for the HM condition predicted by the emotion integration model, and our behavioral data revealed a prediction error mechanism in pain modulation by negative expectations (see Results), we used the prediction error of pain in the HM condition (relative to the MM condition; see Behavioral analysis and computational models section above) as a covariate to examine whether the amygdala and/or ACC activity covaried with pain prediction errors (contrast “HM > MM”). To correct for the number of ROIs, the left amygdala and ACC ROIs were combined into a single mask for the SVC analysis.
Psychophysiological interaction (PPI) analysis
To examine whether the functional interaction between the amygdala and ACC and the functional interaction between the mOFC and the aIC and/or aHPC respectively reflected the negative and positive expectation-associated posterior predicted by the emotion integration model, we performed PPI analyses, a measure of context-dependent connectivity that assesses the degree to which the statistical dependence of time series between a seed region (here, amygdala and mOFC) and other brain regions (ACC for amygdala; aIC and aHPC for mOFC) changes as a function of experimental conditions (Friston et al., 1997). In this analysis, we used the suprathreshold amygdala and mOFC voxels (identified from the above-described first step of our fMRI data analysis) as seed, with the predicted posterior (Eq. 11) as a covariate in SPM (contrast “Attend HM” for negative expectations and contrast “Attend LM” for positive expectations). To control for multiple comparisons during SVCs, the aIC and aHPC ROIs were combined into a single mask for the SVC analysis in the mOFC-seeded PPI analysis in positive expectations. The PPI analysis was carried out using the generalized PPI toolbox (http://www.nitrc.org/projects/gppi) (McLaren et al., 2012). In each participant, physiological variables were generated by extracting the deconvolved times series from each seed region. The PPI regressor representing the interaction between physiological time series and experimental context was formed by multiplying this time series with the task contrast associated with the 10 regressors during painful stimulation. Each participant's first-level PPI contrast images were then entered into a second-level GLM for random-effects group analysis.
ROI definition
To replicate the involvement of the ACC and aIC in pain modulation by stimulus expectancy, we created the ROI for the ACC [a 6 mm radius sphere centered at MNI coordinates (0, 34, 22)] and aIC [a 6 mm radius sphere centered at MNI coordinates (34, 26, 10)] according to our prior study in which stimulus expectancy manipulations were similar to those adopted in the current study (Shih et al., 2019). For emotion-associated neural substrates, we defined (1) the ROI for the bilateral amygdala [two 6 mm radius spheres centered at MNI coordinates (−21, −5, −16) and (22, −4, −15)] according to prior meta-analysis on amygdala response during emotion processing (Sergerie et al., 2008), (2) the mOFC ROI [a 12 mm radius sphere centered at MNI coordinates (6, 46, −15)] according to another meta-analysis on brain activation related to pleasantness (Kuhn and Gallinat, 2012), and (3) the ROI for the right aHPC [a 6 mm radius sphere centered at MNI coordinates (34, −6, −24)] based on our previous study, in which we demonstrated the involvement of this region in positive expectancy effects on pain (Shih et al., 2019). The mOFC ROI was pre-defined as a 12 mm radius (but not 6 mm radius) sphere because it is a larger cortical region compared to the amygdala and aHPC.
For group-level whole-brain analyses, we used Statistical non-Parametric Mapping software (SnPM13; http://warwick.ac.uk/snpm) to control for false-positives caused by multiple testing (Nichols and Holmes, 2002; Eklund et al., 2016). This procedure included 5,000 permutations without variance smoothing, a cluster-forming threshold p < 0.001, and a cluster-level family-wise error (FWE) rate of p < 0.05 was set as the threshold to indicate the significance. Within ROIs, we performed SVC analyses, with an FWE-corrected voxel-wise threshold of p < 0.05. Note that, to accurately present SVC results, only suprathreshold voxels were illustrated in the present study.
Additional behavioral experiment
Although our model comparison analysis using Pre-Test anxiety and pleasantness ratings for the Emoprior and Emolike in Equations 7–10 showed that the emotion integration model contributed to pain modulation by stimulus expectancy, it remains unknown whether the same conclusion can be reached if trial-wise anxiety and pleasantness ratings were used. To confirm this conclusion, we conducted an additional behavioral experiment.
Participants
We recruited additional 22 right-handed participants (11 females; aged 18–27 years, mean ± SD = 21.9 ± 2.3 years) in this behavioral experiment. There was no significant difference in age as compared to those recruited in the fMRI experiment (p = 0.386, unpaired t test).
Experimental protocol
In this behavioral experiment, all experimental procedures and trial composition were the same as the fMRI experiment except for the following. First, participants were not scanned. Second, we added a 7 s emotion rating period (using the same 0–100 VAS for anxiety and pleasantness as described in the test session; 3.5 s for each rating) after the cue period and prior to the painful stimulus (Fig. 1B) in one-third of trials for each trial type during Test1 and Test2 (Fig. 1A). For example, two of six HM trials per regulation condition per session encompassed the 7 s emotion rating period.
Behavioral analysis and computational modeling
All behavioral data analyses and model comparison procedures were the same as the fMRI experiment, except that, in Equations 7 and 8, the Emoprior and Emolike respectively refer to the mean anxiety rating for the high-pain cue (or pleasantness rating for the low-pain cue) and the mean anxiety (or pleasantness) for the medium-pain cue during the 7 s emotion rating period.
Data availability
De-identified behavioral and fMRI data are available on the Open Science Framework (https://osf.io/943df/?view_only=74fe6835c6f44f6c881c74826c9403d4) and NeuroVault (https://neurovault.org/collections/ZCZVVVTW/).
Results
Behavior 1: the presence of emotional components in pain modulation by stimulus expectancy
For Attend trials in the test sessions, by using participants’ pain rating as a dependent variable in a 3 (pain intensity in Attend trials) × 2 (trial in each fMRI session: the first or last trial) repeated-measures ANOVA, there was a significant main effect of pain intensity (all p < 0.0001) but no significant interaction (all p ≥ 0.243). Post hoc comparisons demonstrated no significant difference in pain ratings for HH, MM, and LL conditions between the first and last trial in each fMRI session (all p ≥ 0.541), indicating no significant habituation or sensitization of pain ratings. Our participants exhibited significant negative and positive expectancy effects on pain during both Attend and Regulate conditions (Table 1). For stimulus expectancy effects on pain, a 2 (expectation: negative or positive) × 2 (instruction: Attend or Regulate) repeated-measures ANOVA revealed a significant main effect of instruction (F1,30 = 19.5, p = 0.0001), but no significant interaction (p = 0.871). Post hoc Bonferroni-corrected testing confirmed significant differences in stimulus expectancy effects between Attend and Regulate conditions in both negative expectations (t30 = 3.48, p = 0.003) and positive expectations (t30 = 3.25, p = 0.006). When we used SCRs as the dependent variable (Fig. 3), this two-way repeated-measures ANOVA still showed a significant main effect of instruction (F1,28 = 4.76, p = 0.038) with no significant interaction (p = 0.552), and Post hoc Bonferroni-corrected testing revealed significant differences in SCRs between Attend and Regulate conditions (negative expectation: t28 = 2.86, p = 0.016; positive expectation: t28 = 3.71, p = 0.002). These findings confirm the efficacy of ER, with comparable extent of ER-induced changes in stimulus expectancy effects between negative and positive expectations.
Skin conductance response during pain anticipation. Compared to Attend (Att) trials, the skin conductance responses associated with the high-pain (H) and low-pain (L) cue [relative to the medium-pain (M) cue] were both significantly reduced in Regulate (Reg) trials (p = 0.016 and 0.002, respectively; two-way repeated-measures ANOVA, Bonferroni's post hoc test). Each dot represents data from a participant. Thick lines and error bars indicate mean with SD. *p < 0.05.
Stimulus expectancy effects on pain ratings
If pain expectancy did provoke some emotions that modulated pain processing during stimulation, we expected that this component should meet the following two requirements. First, in Attend trials, the magnitude of this emotional response should be different when participants anticipated high (or low) pain versus medium pain. Second, the magnitude of this emotional response should change after ER, and this change should predict ER-elicited change in stimulus expectancy effects.
For Attend trials in negative expectations, we found that the anxiety level associated with a high-pain cue was significantly higher than that evoked by a medium-pain cue (58.1 ± 22.7 vs 33.0 ± 19.1, p < 0.0001). Compared to Attend trials, both the extent of negative expectancy effect on pain and anticipatory anxiety of the high-pain cue (relative to the medium-pain cue) were significantly reduced in Regulate trials (negative expectancy effect, 14.0 ± 8.4 vs 11.0 ± 7.4, p = 0.004; anxiety, 25.1 ± 16.5 vs 10.0 ± 13.8, p = 0.0004, respectively; Fig. 4A,B), and there existed a significant positive correlation between these two reductions (r = 0.373, p = 0.025, one-tailed; Fig. 4C). These findings suggest a role of anxiety in the pain-enhancing effects of negative expectations.
A role of emotion in pain modulation by stimulus expectancy. A–C, Compared to Attend (Att) trials, both the pain-enhancing effects of negative expectations and anticipatory anxiety associated with the high-pain (H) cue [relative to the medium-pain (M) cue] were significantly reduced in Regulate (Reg) trials (p = 0.004 and 0.0004, respectively; paired t test), and there was a significant intersubject positive correlation between these two reductions (p = 0.025; one-tailed). D–F, Likewise, emotion downregulation significantly reduced both the pain-attenuating effects of positive expectations and anticipatory pleasantness associated with the low-pain (L) cue (relative to the medium-pain cue; p = 0.004 and 0.022, respectively; paired t test), with a significant intersubject positive correlation between these two reductions (p = 0.027; one-tailed). G, Compared to Attend trials, the reduction in left amygdala activation as negative expectations enhanced pain (contrast “HM > MM”) in Regulate trials paralleled the reduction in anticipatory anxiety elicited by a high-pain cue (relative to the medium-pain cue). H, Similar analyses were carried out in positive expectancy effect (contrast “LM > MM”), in which emotion regulation-induced reduction in medial orbitofrontal cortex (mOFC) activation covaried with the reduction in anticipatory pleasantness elicited by a low-pain cue (relative to a medium-pain cue). All depicted clusters in statistical parametric maps survived small-volume corrections with a voxel-wise threshold of p < 0.05 (FWE corrected) and were overlapped on an average structural image. The color bar shows SPM-derived t scores. Each dot represents data from a participant. In A,B,D,E, thick lines and error bars indicate mean with SD. *p < 0.05.
With regard to positive expectations, we found that a low-pain cue evoked a significantly higher level of pleasantness compared to a medium-pain cue in Attend trials (47.0 ± 32.5 vs 28.5 ± 21.3, p = 0.0004). ER significantly reduced both the extent of positive expectancy effect (18.3 ± 8.1 vs 15.6 ± 8.0, p = 0.004; Fig. 4D) and anticipatory pleasantness elicited by the low-pain cue (relative to the medium-pain cue; 18.5 ± 25.9 vs −0.65 ± 25.9, p = 0.022; Fig. 4E), with these two reductions positively correlated with each other (r = 0.348, p = 0.027, one-tailed; Fig. 4F). Notably, ER-elicited reduction in pleasantness ratings was significantly associated with an increase in anxiety ratings (r = 0.418, p = 0.019), and this increase in anxiety marginally predicted ER-elicited reduction in positive expectancy effect (r = 0.272, p = 0.070, one-tailed). Compared with the reduction in pleasantness ratings, the magnitude of ER-elicited increase in anxiety ratings was significantly smaller (p = 0.015). These findings suggest a major role of pleasantness, as well as a minor role of anxiety, in the pain-attenuating effects of positive expectations.
In terms of the relationship between the emotional components associated with negative and positive expectations, ER-elicited reduction in stimulus expectancy effects did not show a significant relationship between these two types of expectations (p = 0.689), which implies that pain modulation by negative and positive expectations might engage dissimilar emotional components.
fMRI 1: brain structures encoding the emotional component of pain modulation by stimulus expectancy
During the anticipation period, ER induced significant activation in the lateral prefrontal region (inferior frontal gyrus; contrast “Regulate > Attend” pooled over the three pain-predictive cues; whole-brain corrected; Table 2), a key region involved in ER (Delgado et al., 2008b; Kohn et al., 2014). During painful stimulation, typical brain regions tracking pain intensity, such as the primary somatosensory cortex, were activated (whole-brain corrected; Table 3). By pooling both Regulate and Attend conditions together, we replicated our previous research showing that the ACC exhibited increase responsivity when negative expectations enhanced pain [contrast “HM > MM”; peak MNI = 4/38/20, t27 = 3.25, k = 5, p = 0.042, voxel-wise small-volume family-wise error corrected (SVC-FWE); Fig. 5A], and the right aIC showed increase responsivity when positive expectations alleviated pain (contrast “LM > MM”; peak MNI = 38/26/6, t30 = 3.76, k = 7, p = 0.013, SVC-FWE; Fig. 5B; Shih et al., 2019). BOLD activation within these ACC and aIC clusters was not significantly different between Attend and Regulate trials (both p ≥ 0.232).
Brain activation associated with pain modulation by stimulus expectancy. A, The anterior cingulate cortex (ACC) exhibited increased activation as negative expectations enhanced perceived pain (contrast “HM > MM” pooled over Attend and Regulate conditions). B, The right anterior insular cortex (aIC) showed increased activation as positive expectations reduced perceived pain (contrast “LM > MM” pooled over Attend and Regulate conditions). All depicted clusters in statistical parametric maps survived small-volume corrections with a voxel-wise threshold of p < 0.05 (FWE corrected) and were overlapped on an average structural image. The color bar shows SPM-derived t scores. In dot plots, each dot represents data extracted from the suprathreshold cluster in a participant, and thick lines and error bars indicate mean with SD.
Whole-brain activations associated with emotion regulation during the cue period
Whole-brain activations during painful stimulation
Given changes in emotions paralleled changes in pain modulation by stimulus expectancy (Fig. 4C,F), if pain expectancy-elicited emotion was represented by a brain region that modulated pain processing, it would be plausible to anticipate that its responsivity should track ER-elicited change in anxiety (or pleasantness) as pain modulation declined. On the neural level, we hypothesized that the amygdala would represent the emotional component of negative expectations (i.e., anxiety), given amygdala's crucial role in processing anxiety (Tovote et al., 2015). As the emotional responses involved in positive expectations encompassed both pleasantness and anxiety, and a low-pain cue signaled low threat, we tested whether these emotional responses would be reflected in (1) the mOFC, a region implicated in processing both pleasantness (Kuhn and Gallinat, 2012) and anxiety (Adhikari et al., 2010, 2011; Padilla-Coreano et al., 2016), and (2) the aHPC, a region involved in pain modulation by positive expectations (Shih et al., 2019) and threat reduction in the presence of relatively safe information (Meyer et al., 2019).
As hypothesized, during painful stimulation, ER-induced reduction in left amygdala activation covaried with anxiety reduction [contrast “Attend (HM > MM) > Regulate (HM > MM)”; peak MNI = −22/−4/−12, t26 = 3.52, k = 1, p = 0.043, SVC-FWE; Fig. 4G], whereas a reduction in mOFC activation was positively correlated with pleasantness reduction [contrast “Attend (LM > MM) > Regulate (LM > MM)”; peak MNI = 12/48/−12, t29 = 4.79, k = 4, p = 0.008, SVC-FWE; Fig. 4H]. There was no significant correlation between BOLD activations within the amygdala and mOFC (p = 0.246).
Behavior 2: emotional integration of expectations and sensory inputs in pain modulation by stimulus expectancy
After identifying the emotional component in stimulus expectancy effects via comparing the Regulate against Attend condition, we next investigated how this emotional component modulated pain perception in the Attend condition. Following a prior study (Grahl et al., 2018) that adopted a Bayesian integration framework (Eq. 1 in Materials and Methods), we examined the precision integration model, in which pain expectations (prior; pain ratings of HH or LL trials during the conditioning session) and incoming sensory inputs (likelihood; pain ratings of Attend MM trials during the test session) were respectively weighted by the relative precision (i.e., the inverse variance of pain ratings) of prior (Wprior; see Eq. 2 in Materials and Methods) and likelihood (Wlike; Eq. 3 in Materials and Methods) to account for pain modulation by stimulus expectancy (Fig. 2A,B). Alternatively, we propose another emotion integration model, in which expectations and sensory inputs were both weighted by their respective emotion (anxiety ratings for negative expectations, and pleasantness and anxiety ratings for positive expectations) to form the experience of pain (Fig. 2C,D). We anticipated that the relative emotion regarding the prior and likelihood (Wprior_Emo and Wlike_Emo; see Eqs. 7–10 in Materials and Methods) would modulate their influence on the estimation of the posterior (Eq. 11 in Materials and Methods).
For the precision integration model, we found that neither the relative precision of the prior (Eq. 2) nor the hypothetical stimulus expectancy effect (Eqs. 5, 6) predicted the observed stimulus expectancy effect by negative and positive expectations (all p ≥ 0.561). Of note, the hypothetical stimulus expectancy effect based on the emotion integration model (Eqs. 5, 6) positively predicted observed stimulus expectancy effects. For anxiety in negative expectations, the observed negative expectancy effect covaried with the hypothetical negative expectancy effect (r = 0.563, p = 0.0009; one-tailed; Fig. 6A). For positive expectations, consistent with the role of both pleasantness and anxiety in pain modulation, the observed positive expectancy effect was significantly correlated with the hypothetical value based on either pleasantness ratings (r = 0.466, p = 0.004; one-tailed; Fig. 6B) or anxiety ratings (r = 0.523, p = 0.001; one-tailed). Here, for both prior and likelihood in these two integration models, their relative emotion (Wprior_Emo and Wlike_Emo) was independent on their relative precision (Wprior and Wlike; all p ≥ 0.396).
Behavioral and fMRI findings based on emotion integration model. A, The hypothetical negative expectancy effect estimated by anxiety-associated emotion integration model (Fig. 2C and Eq. 5 in Materials and Methods) positively predicted observed negative expectancy effect. B, The pleasantness-associated hypothetical positive expectancy effect (Fig. 2D and see Eq. 6 in Materials and Methods) positively predicted observed positive expectancy effect. C, For negative expectations, the anticipatory anxiety related to pain expectations (prior) positively predicted observed negative expectancy effect. D, For positive expectations, the relative pleasantness between pain expectations and sensory inputs (Wprior_Pleasantness; Eq. 7 in Materials and Methods) positively predicted observed positive expectancy effect. E, For negative expectations, by performing a psychophysiological interaction (PPI) analysis (contrast “Attend HM”), the strength of functional connectivity from the left amygdala to anterior cingulate cortex (ACC) covaried with individual hypothetical pain perception
To compare these two integration models, we then performed random-effects Bayesian model selection (Stephan et al., 2009; Rigoux et al., 2014) to estimate the overall posterior model probability across participants (Fig. 2E–H). For negative expectations, we found that 8 participants showed a BF greater than three in favor of the emotion integration model over the precision integration model, in contrast to two participants showing a BF greater than three in favor of the precision integration model over the emotion integration model (Fig. 2E). The random-effects model comparison analysis yielded an expected posterior model probability of 0.909 (exceedance probability > 0.999) for the emotion integration model, compared to 0.091 (exceedance probability < 0.001) for the precision integration model (Fig. 2F). Likewise, for positive expectations, 14 participants showed a BF greater than three in favor of the emotion (pleasantness) integration model over the precision integration model, but only four participants showed a BF greater than three in favor of the precision integration model over the emotion (pleasantness) integration model (Fig. 2G). The expected posterior model probability was 0.918 (exceedance probability > 0.999) for the emotion integration model and 0.082 (exceedance probability < 0.001) for the precision integration model (Fig. 2H). These findings strongly favored the emotion integration model over the precision integration model.
To verify this conclusion, we conducted an additional behavioral experiment, in which we collected trial-wise emotion rating to estimate model parameters and also perform model comparisons in the Attend condition (see Additional behavioral experiment section in Materials and Methods). Consistent with our fMRI experiment, this behavioral experiment yielded very similar model comparison results: for negative expectations, the expected posterior model probability was 0.906 (exceedance probability > 0.999) for the emotion integration model and 0.094 (exceedance probability < 0.001) for the precision integration model; for positive expectations, the expected posterior model probability was 0.913 (exceedance probability > 0.999) for the emotion integration model and 0.087 (exceedance probability < 0.001) for the precision integration model.
Taken together, these behavioral findings confirm that the emotion (but not precision) associated with expectations and sensory inputs contributed to pain modulation by stimulus expectancy. To further examine whether there existed differences in the role of emotion between negative and positive expectations, we then checked the relationships between the observed stimulus expectancy effects and the emotion responses associated with expectations and sensory inputs in our fMRI experiment. In negative expectations, we found that the anxiety rating for pain expectations provoked by a high-pain cue significantly predicted the observed negative expectancy effect on pain (r = 0.489, p = 0.004; one-tailed; Fig. 6C), but the anxiety rating for perceived sensory inputs and the relative anxiety between expectations and sensory inputs did not (both p ≥ 0.153; one-tailed). By contrast, it was the relative pleasantness between pain expectations and sensory inputs (Wprior_Pleasantness) that significantly predicted the observed positive expectancy effect on pain (r = 0.330, p = 0.035; one-tailed; Fig. 6D), but not the pleasantness rating for prior stimulus experience or for perceived sensory inputs (both p ≥ 0.185; one-tailed). Again, the relative anxiety between expectations and sensory inputs (Wprior_Anxiety) also predicted the observed positive expectancy effect (r = 0.357, p = 0.024; one-tailed).
fMRI 2: neural basis underlying emotional integration in pain modulation by stimulus expectancy
Following the behavioral results that supported the emotion integration model, we then adopted two approaches to investigate how emotions modulated the influence of stimulus expectancy on pain at the neural level. In the first approach, given ACC and aIC responses underlay stimulus expectancy effects on pain (Fig. 5; Shih et al., 2019), we asked whether amygdala and mOFC (Fig. 4G,H) were functionally linked to them (i.e., functional interactions between amygdala and ACC for negative expectations and between mOFC and aIC for positive expectations) during this integration process. This would reinforce the notion that the emotional component of pain expectations contributed to stimulus expectancy effects on pain. For positive expectations, we previously showed that the extent of pain modulation was inversely correlated with anxiety and engaged the hippocampus as well, suggesting the involvement of aHPC in pain modulation by positive expectations (Shih et al., 2019). Therefore, in addition to the aIC, we also tested whether the mOFC (Fig. 4H) was functionally linked to the aHPC to support emotional integration of expectations and sensory inputs as positive expectations attenuated pain.
For negative expectations, we found that the functional connectivity between the left amygdala and ACC covaried with anxiety-associated posterior for the HM condition predicted by the emotion integration model (Eq. 11; contrast “Attend HM”; peak MNI = −4/36/18, t26 = 3.59, k = 4, p = 0.018, SVC-FWE; Fig. 6E). Regarding positive expectations, the functional connectivity from the mOFC to aHPC covaried with the pleasantness-associated posterior for the LM condition (Eq. 11; contrast “Attend LM”; peak MNI = 38/−6/−22, t29 = 4.46, k = 9, p = 0.005, SVC-FWE; Fig. 6F). Again, the connectivity strength extracted from the suprathreshold aHPC cluster – which reflected individual mOFC-aHPC coupling when positive expectations reduced pain – also paralleled the anxiety-associated posterior for the LM condition (Eq. 11; contrast “Attend LM”; r = 0.334, p = 0.033).
Behavior 3 and fMRI 3: prediction error mechanism in pain modulation by stimulus expectancy
In the second approach, we investigated whether the effect of emotion on expectancy-associated pain modulation would be accompanied by an operation of prediction error mechanism. The rationale behind this approach is that prediction error signals have been demonstrated to be modulated by emotion (Watanabe et al., 2013), and suppression of these signals has been thought to underlie the maintenance of expectancy modulation on pain (Schenk et al., 2017).
For the magnitude of pain prediction errors, a 2 (negative or positive) × 2 (Attend or Regulate) repeated-measures ANOVA showed a significant interaction (F1,27 = 12.93, p = 0.001). Post hoc Bonferroni-corrected testing confirmed a significant difference in pain prediction errors between Attend and Regulate conditions in negative expectations (t27 = 3.060, p = 0.010) but not positive expectations (t27 = 2.025, p = 0.106; Fig. 7A). For the expected pain intensity, another 2 (negative or positive) × 3 (Pre-Test, Attend trials, or Regulate trials) repeated-measures ANOVA also showed a significant interaction (F2,54 = 23.51, p < 0.0001), with post hoc Bonferroni-corrected testing revealing that only the pairwise comparisons between Regulate trials and Attend trials (t54 = 6.089, p < 0.0001) and between Regulate trials and Pre-Test (t54 = 7.592, p < 0.0001) in negative (but not positive) expectations reached statistical significance (Fig. 7B).
Prediction error mechanism in pain modulation by stimulus expectancy. A, For negative expectations, the magnitude of pain prediction error in the HM condition during scanning was significantly reduced in Regulate (Reg) trials compared to Attend (Att) trials (p = 0.010; two-way repeated-measures ANOVA, Bonferroni's post hoc test). For positive expectations, this comparison did not reach statistical significance (p = 0.106). B, For negative expectations, compared to the pre-scanning rating (Pre-Test; Fig. 1A), the Attend trials during scanning (Post-Test1 and Post-Test2) did not entail a significant difference in participants’ expected pain intensity rating regarding the high-pain (H) cue (p = 0.833; two-way repeated-measures ANOVA, Bonferroni's post hoc test), but this rating was significantly lower in Regulate trials relative to Attend trials during scanning (p < 0.0001; repeated-measures ANOVA, Bonferroni's post hoc test). For positive expectations, none of these two pairwise comparisons regarding the low-pain (L) cue yielded significant results (both p > 0.999; two-way repeated-measures ANOVA, Bonferroni's post hoc test). C, Emotion regulation-induced changes in expected pain intensity for the high-pain cue covaried with corresponding changes in anticipatory anxiety (p = 0.042). D, Activation within the anterior cingulate cortex (ACC) showed a significant intersubject positive correlation with the prediction error of pain in the HM condition (relative to the MM condition; contrast “HM > MM”) in Regulate trials. In A and B, thick lines and error bars indicate mean with SD. All depicted clusters in statistical parametric maps survived small-volume corrections with a voxel-wise threshold of p < 0.05 (FWE corrected) and were overlapped on an average structural image. The color bar shows SPM-derived t scores. *p < 0.05.
These findings suggest that, as negative expectations modulated pain perception, downregulating negative expectation-associated anxiety significantly reduced both the magnitude of pain prediction errors (i.e., resulting in a smaller mismatch between expected and perceived pain) and the expected high pain. In other words, the pain-enhancing effect of negative expectations in the Attend condition might be subserved by the suppression of prediction error signals as well as impaired updating of expected pain. Importantly, this notion is supported by the presence of a significant positive correlation between the extent of downregulated anxiety by ER and the extent of decrease in expected pain intensity in Regulate trials (relative to Attend trials; r = 0.388, p = 0.042; Fig. 7C); that is, participants’ expected high pain dropped with decreasing anxiety, indicating that negative expectation-associated anxiety was indeed linked to the updating of negative expectations for pain. Along with these behavioral findings, we found that activation within the ACC – which coactivated with the amygdala to integrate the expected pain with sensory inputs (Fig. 6E) – covaried with the pain prediction error in the HM condition (relative to the MM condition) in Regulate trials [contrast “Regulate (HM > MM)”; peak MNI = 0/36/26, t26 = 4.18, k = 6, p = 0.012, SVC-FWE; Fig. 7D] but not in Attend trials [contrast “Attend (HM > MM)”; p = 0.683].
Discussion
Herein, we report that negative expectation-elicited anxiety not only influenced the integration of prior expectations with sensory inputs, but dampened aversive prediction error signals and updating of negative expectations. Positive expectation-elicited pleasantness (and decrease in anxiety) modulated the integration of prior expectations with sensory inputs to support the pain-attenuating effects of prior expectations. The finding that the reduction in both negatively (anxiety) and positively (pleasantness) valenced emotions paralleled SCR reduction suggests that SCRs reflect the autonomic arousal (Mauss and Robinson, 2009) induced by pain expectancy. Given the lateral prefrontal region activated during ER has been reported as belonging to attentional networks (Ochsner and Gross, 2005), and attentional distraction downregulates both negative and positive emotional responses (Kanske et al., 2011; Schonfelder et al., 2014), we speculate that the ER strategy we used may reduce anticipatory emotions and associated SCRs via distraction (Delgado et al., 2008a).
Behaviorally, we demonstrated that manipulating participants’ anticipatory anxiety biased pain modulation by negative as well as positive expectations. This is consistent with previous studies showing that an increase (or decrease) in anxiety before the receipt of painful stimulation enhances (or reduces) pain perception (Shih et al., 2019). Along with anxiety, when positive expectations reduced pain, lowering participants’ pleasant feeling also reduced positive expectation-associated pain-attenuating effects. This resonates with previous studies showing that positive expectations for pain elicit reward-related neural responses (Seymour et al., 2005; Scott et al., 2007) and make aversive stimuli become relatively more pleasant (Ellingsen et al., 2013). Interestingly, ER-elicited reduction in pleasantness paralleled an increase in anxiety, whose magnitude was smaller than pleasantness changes. This finding conforms to the opponent process framework, in which the reduction in one sensation is accompanied by a smaller sensation of the opposite valence (Solomon and Corbit, 1974; Seymour et al., 2005). Combined with the finding that pain modulation by positive expectations was significantly predicted by changes in pleasantness but marginally predicted by changes in anxiety, we think that pleasantness constitutes the main emotional component as pain perception is attenuated.
Neurally, we found that amygdala activation reflected negative expectation-associated anxiety. This is in line with the implication of the amygdala in emotional arousal and anxiety (Tovote et al., 2015). Combined with previous reports that amygdala mediates pain enhancement by negative expectations (Atlas et al., 2010; Schmid et al., 2015), we think that amygdala subserves the pain-enhancing effect of anxiety evoked by negative expectations. Regarding positive expectation, we demonstrated that mOFC represented expectation-associated pleasantness. This finding aligns with the involvement of mOFC when pain is paired with positive hedonics (Ellingsen et al., 2013; Leknes et al., 2013). Given ER-induced pleasantness changes predicted changes in pain modulation by positive expectations, we believe that the increased pleasantness evoked by positive expectations recruits the mOFC to alleviate pain. Because ER-elicited change in pain modulation by negative expectations was uncorrelated with that by positive expectations, and there was no significant relationship between amygdala and mOFC activations, our data suggest that negative and positive expectations engage dissimilar emotions and anatomical substrates to modulate pain. Although debated, they also support the notion that negative and positive emotions are represented by separate and functionally independent neural systems (Phan et al., 2002; Vytal and Hamann, 2010; Lindquist et al., 2016).
Like treatment expectancy (Grahl et al., 2018), we demonstrated that pain modulation by stimulus expectancy depended on the integration of prior (expectations) and novel (sensory inputs) pain experiences. Of note, this integration was modulated by above-described anxiety and pleasantness, which confirms the pivotal role of emotions in the context of stimulus expectancy. Interestingly, participants’ anticipatory anxiety itself predicted pain modulation by negative expectations, whereas the relative pleasantness (as well as anxiety) between expectations and sensory inputs determined pain modulation by positive expectations. In our experimental context, the high- (or low-) pain cue was the worst (or best) cue outcome, produced a high- (or low-) threat state, and thus engendered increased anxiety (or pleasantness) relative to the medium-pain cue. When negative expectations enhanced pain (i.e., HM trials), the perceived threat level of experiencing moderate pain did not surpass that of anticipating high pain. Under such circumstances, it is plausible that participants’ processing resources remained prioritized to expectations (i.e., high threat) (Legrain et al., 2009; Pessoa, 2009), which explains why expectation-elicited anxiety directly predicted pain modulation. By contrast, when positive expectations attenuated pain (i.e., LM trials), the perceived threat level of experiencing moderate pain was higher than that of expectations. In such a situation, our findings suggest a role of sensory input-associated emotions in predicting pain modulation. This phenomenon is coherent with previous observation that the effect of sensory inputs on pain perception increases as expectations become more positive (Lim et al., 2020).
Parallel to these behavioral findings, our fMRI data revealed that amygdala-ACC coupling integrated negative expectations and sensory inputs to underlie pain perception (i.e., posterior). The ACC, which has extensive connections with the amygdala (Devinsky et al., 1995), has been assumed to receive threat-related information from the amygdala. Interactions between the amygdala and ACC are not only implicated in fear and anxiety expression (Tovote et al., 2015), but responsible for sustained threat responses (Levy and Schiller, 2021). Regarding positive expectations, we found that pain attenuation related to the increased pleasantness (or reduced anxiety) mainly involved mOFC-aHPC coupling. The mOFC receives direct projections from the aHPC (Parent et al., 2010) and tracks contextual information to suppress threat responding via safety (Schiller et al., 2008). The aHPC controls threat-elicited responses in a low-threat state (Sotres-Bayon et al., 2012; Meyer et al., 2019) – such as the positive expectation condition in the present study. Synchronization between the mOFC and aHPC subserves anxiety-related responses (Adhikari et al., 2010; Padilla-Coreano et al., 2016).
Along with the emotion integration model, we further showed that, when negative expectation-associated anxiety was downregulated, both the prediction error of pain and expected pain intensity reduced. The decrease in anxiety was monotonically related to the decrease in expected pain, and the ACC represented pain prediction errors. Given prediction errors update expectations (Schultz et al., 1997; Sutton and Barto, 1998), these findings imply that anxiety suppresses prediction error signals in the ACC and associated updating of pain predictions as negative expectations enhance pain (i.e., in the Attend condition). This is consistent with previous observations that high-anxiety individuals have not only reduced prediction error representation in the ACC (White et al., 2017), but deranged updating of expectations in threatening contexts (Browning et al., 2015) that is accompanied by aberrant ACC responsivity (Piray et al., 2019). As the ACC is a key node that allocates processing resources to threat, and processing resources devoted to high-threat stimuli would detract from the resources available to other cognitive activity (Pessoa, 2009), we propose that the anxiety responses provoked by anticipating imminent high pain suppress correct adjustment of pain expectations via the impairment of ACC-associated prediction error signaling, making negative expectancy modulation on pain resistant to extinction. Given downregulation of positive expectation-associated pleasantness was accompanied by an increase in anxiety, which presumably suppressed prediction error signals and updating of expectations, future studies are required to clarify whether pleasantness indeed influences adequate updating of positive expectations for pain.
It remains debated whether pain modulation by stimulus and treatment expectancy involves dissimilar mechanisms (Atlas and Wager, 2014). In typical treatment expectancy paradigm such as placebo hypoalgesia, participants were explicitly informed of why a procedure (e.g., placebo cream application) could reduce pain, and participants’ expectation is usually fixed during an experimental session (Wager et al., 2004; Grahl et al., 2018). In this context, tonic modulation on pain is required, with the precision of prior experiences acquired in a preceding conditioning session playing a major role in determining expectancy effects (Grahl et al., 2018). By contrast, in stimulus expectancy, participants acquired cue-stimulus contingencies mainly through trial-by-trial learning during conditioning, and their expectations change in a trial-wise manner during the test session. In such an ever-changing environment, our Bayesian model comparisons in both the fMRI and additional behavioral experiments demonstrated the key role of emotions (relative to precision) in pain modulation, which is consistent with the concept that emotional information is prioritized in processing when subjects encounter threatening stimuli (LeDoux, 2012). Taking into account the overlapping neural substrates (i.e., ACC, aIC, amygdala, and mOFC) engaged in pain modulation by both stimulus expectancy (the current study) and treatment expectancy (Bingel et al., 2006; Wager et al., 2007; Ellingsen et al., 2013; Tinnermann et al., 2017), we believe that the distinction between stimulus and treatment expectancy effects may reside in the underlying computational mechanisms (precision- vs emotion-modulated integration).
Conclusion
In conclusion, we demonstrate a key role of emotion processing in pain modulation by stimulus expectancy. Our findings not only add to the growing studies investigating the computational mechanisms behind expectancy modulation on pain (Yoshida et al., 2013; Grahl et al., 2018; Jepma et al., 2018), but help explain why subjects with chronic pain, who frequently experience comorbid mood disorders (Wiech and Tracey, 2009), exhibit dysregulated expectations for pain (Kaptchuk et al., 2020), pain perception, and pain-evoked neural responses (Bushnell et al., 2013).
Footnotes
This work was supported by the Ministry of Science and Technology of Taiwan (MOST 108-2410-H-002-110-MY2, 110-2314-B-002-165-MY3, and 111-2314-B-002-255-MY3). We appreciate the support from the Imaging Center for Integrated Body, Mind and Culture Research in National Taiwan University.
The authors declare no competing financial interests.
- Correspondence should be addressed to Ming-Tsung Tseng at mingtsungtseng{at}ntu.edu.tw.
