The temporal characteristics of expectations and prediction errors in pain

In the context of a generative model, such as predictive coding, pain perception can be construed as the integration of expectation and nociceptive input with their difference denoted as a prediction error. In a previous neuroimaging study (Geuter et al., 2017) we observed an important role of the insula in such a model, but could not establish its temporal aspects. Here we employed electroencephalography to investigate neural representations of predictions and prediction errors in heat pain processing. Our data show that alpha-to-beta activity was associated with pain expectation, followed by gamma band activity associated with absolute prediction errors. Source reconstruction revealed the insula as a common region for both effects. This temporal sequence of expectation related alpha-to-beta activity, followed by prediction error associated gamma activity in the insula, provides a possible mechanisms for the temporal integrating of pain predictions and prediction errors in the context of a generative model.


29
It has been shown that physically identical nociceptive input, can lead to variable 30 sensations of pain, depending on contextual factors (Tracey and Mantyh, 2007). In 31 particular, attention, reappraisal and expectation are core mechanisms that influence how 32 nociception leads to pain (Wiech et al., 2008). A clinically important example of how 33 expectations can shape pain processing is placebo hypoalgesia: pain relief mediated by 34 expectation and experience -in the absence of active treatment (Petrovic et Anchisi and Zanon, 2015). 37 In the context of a generative model of pain, it has been proposed that pain perception can 38 be seen as the consequence of an integration of expectations with nociception (Büchel et 39 al., 2014;Wiech, 2016;Ongaro and Kaptchuk, 2019). In this framework, expectations are 40 integrated with incoming nociceptive information and both are weighted by their relative 41 precision (Grahl et al., 2018) to form a pain percept. This can be seen in analogy to ideas 42 in multisensory integration (Ernst and Banks, 2002). Expectations or predictions and 43 resulting prediction errors also play a key role in generative models such as predictive 44 coding (Huang and Rao, 2011). In essence, this framework assumes that neuronal 45 assemblies implement perception and learning by constantly matching incoming sensory 46 data with the top-down predictions of an internal or generative model (Knill and Pouget,47 2004; Huang and Rao, 2011;Clark, 2013). Basically, minimizing prediction errors allows 48 systems to resist their tendency to disorder by the creation of models with better 49 predictions regarding the sensory environment, leading to a more efficient encoding of 50 information (Friston, 2010). However, in fMRI studies predictions and prediction errors cannot be temporally 60 dissociated due to the low temporal resolution of the method. To investigate this further, 61 we conducted a cue based pain experiment using EEG to achieve high temporal and 62 spectral resolution of predictions and prediction error processes in the context of pain. 63 In this experiment (n=29) we employed contact heat stimuli with three different 64 intensities (warm, mildly painful, painful), preceded by a visual cue indicating the 65 upcoming intensity (Figure 1). To study modality effects we also presented aversive 66 pictures with 3 levels of aversiveness. To generate prediction errors, the modality 67 (picture or pain) was correctly cued only in 70% of all trials and stimulus intensities were 68 correctly cued only in 60% of all trials. We then investigated oscillatory activity related to 69 stimulus intensity, expectation and prediction errors (  that prediction error signals should be related to higher frequencies (e.g. gamma band) 87 than prediction signals (Todorovic et al., 2011;Arnal and Giraud, 2012). Finally, based on 88 our previous fMRI results (Geuter et al., 2017;Fazeli and Büchel, 2018) we expect these 89 effects to include the insular cortex and in particular the anterior insula. 90

91
Behavioral data -pain ratings 92 Participants experienced aversive heat or saw picture stimuli which were 93 probabilistically cued in terms of modality and intensity, evoking an expectation of 94 modality and intensity. The subsequently applied stimuli were then rated on a visual 95 analog scale (VAS) from 1-4. Our primary behavioral question was whether ratings are 96 influenced by the experimental manipulation of stimulus intensity, expectation and 97 absolute prediction errors. 98 To validate our intensity manipulation for heat pain and to verify the discriminability 99 between different levels of pain, we first tested for the main effect of stimulus intensity 100

107
To evaluate the main effects of stimulus intensity, expectation and absolute prediction 108 error, we employed a repeated measures ANOVA of the behavioral data, which revealed 109 significant effects for the main effect of stimulus intensity, i.e. the three levels of pain 110 (F(2,28) = 743.97, p < .001). Also, the main effect for expectation on pain ratings was 111 significant (F(2,28) = 38.53, p = < .001) (Table 1), indicating an influence of the cued 112 intensity on behavioral pain ratings. The absolute difference between the cued intensity 113 and the actual stimulus intensity (i. e. prediction error), did not yield a significant effect 114 on pain ratings (F(2,28) = 2.87, p = .10). The results regarding the aversive pictures are 115 summarized in Table 1  shows both an effect of stimulus intensity (increase from blue to green to red), but also an effect 123 of expectation (low to medium to high expectation).

125
In a first EEG analysis, we tested for a main effect of the intensity of the nociceptive input 126 in the context of a correctly cued modality (i.e. pain was expected and received). In order 127 to do so, we performed a repeated measures ANOVA on the time-frequency 128 representation of the EEG data on low frequencies (1-30Hz) and high frequencies (31-129 100Hz) separately after pain onset using a cluster correction criterion to address the This sample was found at 1250ms and 13Hz and had a maximum at channel C4. 142 Also in the low frequency range (1-30Hz), the positive cluster included samples in a time 143 range from 150 to 1250ms after pain onset in a frequency range from 1 to 7Hz at central 144 electrode sites (p = .038). The highest parametric F-Value from the repeated measures 145 ANOVA was F(2,28) = -40.87, (p < .001). This sample was found at 550ms and 1Hz and 146 had a maximum at channel Cz. 147 In the high frequency range (31-100Hz) representing Gamma activity one positive cluster 148 was significant ( Figure 5). This cluster included samples in a time range from 500 to 149 1600ms after pain onset and frequencies from 37 to 100Hz, predominately at contra-150 lateral central electrode sites. The highest parametric F-value within this cluster obtained 151 from the repeated measures ANOVA was F(2,28) = 21.30, (p < .001). This sample was 152 found at 600ms and 60Hz and had a maximum at channel FCz. 153 In conclusion, these results indicate that a higher intensity of the nociceptive input is 154 associated with increased theta and gamma band power and a negative relationship of 155 alpha-to-beta band power and the intensity of the nociceptive input. Using a cued pain paradigm with three different pain intensities, our data showed a clear 225 discriminability of different levels of pain based on behavioral ratings and EEG time-226 frequency patterns. Specifically, we observed several clusters of activity to be associated 227 with the intensity of thermal stimulation in the theta, beta and gamma band. Furthermore, 228 behavioral data clearly indicated a positive influence of cued intensity on pain perception. 229 In addition, our results provide evidence for temporally and spectrally separable clusters 230 of oscillatory activity associated with expectation and prediction errors for pain. 231 Specifically, an early (350-600ms) low frequency (1-30Hz) cluster was related to 232 expectation i.e. cued intensity. In contrast, a later occurring cluster at higher frequencies 233 (31-100Hz) was related to prediction errors.  Figure 4b). In addition we observed a significant suppression of alpha-to-242 beta activity which, given the above mentioned delays of our painful stimuli, is in line with 243 the reported beta suppression in previous EEG studies on pain (Ploner et al., 2006). contextual processes, such as expectations, which is associated with alpha/beta 276 oscillations and alpha/beta synchrony across brain areas. Previous studies have started 277 to examine the spectral properties of mechanisms related to generative models of pain 278 perception. In particular, a previous MEG study reported that alpha suppression in the 279 anterior insula is related mainly to pain expectation in a paradigm in which painful stimuli 280 were interleaved with non-painful stimuli (Franciotti et al., 2009). This was interpreted 281 as a preparatory mechanism for an upcoming painful stimulus. In a related study, alpha 282 desynchronization in the context of predictable painful stimuli, has been discussed as a 283 possible neural correlate of attentional preparatory processes (Babiloni et al., 2003). 284 Expectation is also a crucial ingredient of placebo analgesia and nocebo hyperalgesia. A 285 previous study reported that resting state alpha band activity was also linked to the 286 expectation of pain modulation (analgesia) in a placebo paradigm (Huneke et al., 2013). 287 Source localization linked this effect to the insula, the medial frontal and prefrontal cortex. 288 With respect to negative expectations, it has been shown that pain modulation due to 289 nocebo expectation is associated with enhanced alpha activity (Albu and Meagher, 2016). Although these studies revealed expectation and prediction error effects in the insula, 302 they were unable to provide deeper mechanistic insights into the temporal orchestration 303 of these effects due to the limited temporal resolution of the hemodynamic response 304 function as measured in fMRI. 305 Importantly, here we can show that although both effects are co-localized in the insula 306 ( Figure 8b), there is a precise temporal and spectral organization of both effects within 307 the insula. Initially, low frequency activity is linked to expectation effects which after 308 integration with nociceptive information can lead to a prediction error expressed by 309 activity in the gamma band. This is further supported by the estimated topographies of 310 the effects. Main effects of stimulus intensity, expectation and absolute prediction errors 311 are all represented by a parietal-to-frontal pattern of activation, suggesting a similar 312 associative network to be at play (see Figure 9 for a schematic summary of the results). To unravel, the temporal aspects of expectations and prediction errors, this study has 323 been designed in close analogy to a previous fMRI study and we decided to use the same 324 experimental paradigm (Fazeli and Büchel, 2018). We therefore decided to also keep the 325 sample characteristics similar and restricted the sample to male participants, which 326 means that we cannot generalize our results to the population. However, our study agrees 327 with findings of a previous study using a similar design (Geuter et al., 2017) and which 328 tested male and female participants. Future studies should investigate samples including 329 female participants. This would also allow to investigate sex effects with respect to 330 expectation and prediction error effects in pain. 331 Although EEG has an excellent temporal resolution, its spatial resolution is limited. 332 Consequently, the source reconstruction using LORETA is of low spatial resolution 333 Each trial began with the presentation of the cue for 500ms as an indicator for the 384 modality and intensity of the subsequently presented stimulus. The modality was 385 correctly cued in 70% of all trials by the color of the triangle. In 60% of all trials, the 386 stimulus intensity was correctly indicated by the digit within the triangle (see Fig. 1b for 387 an overview of all cue contingencies). 388 Before the presentation of the stimulus, there was a blank period with a variable time 389 frame between 1000ms and 1400ms. Then, the visual or thermal stimulus was presented 390 for a duration of two seconds. The visual stimulus was centered on the screen and allowed 391 the participant to perceive the stimulus without eye movements. Right after the 392 termination of the stimulus, subjects were asked to rate the aversiveness of the stimulus 393 on a four point rating scale, where 1 was labeled as "neutral" and 4 was labeled as "very 394 strong". Ratings were performed using a response box operated with the right hand (see 395 Figure 1 for a visualization of the trial structure). 396 In addition, four catch trials were included in each block. Subjects were asked to report 397 the preceding cue in terms of their information content of the modality and intensity 398 within 8s and no stimulation was given in these trials. The data was detrended and baseline correction was applied over the whole time range 432 of each trial. The data was band-pass filtered at 1-100Hz, Butterworth, 4th order. EEG 433 epochs were then visually inspected and trials contaminated by artifacts due to gross 434 movements or technical artifacts were removed. Subsequently, trials contaminated by 435 eye-blinks and movements were corrected using an independent component analysis 436 (ICA) algorithm (Makeig et al., 1996;Jung et al., 2000). In all datasets, individual eye 437 movements, showing a large EOG channel contribution and a frontal scalp distribution, 438 were clearly seen in the removed independent components. Back-projection of the 439 remaining non-artifactual components resulted in corrected data. Subsequently, the data 440 was re-referenced to a common average of all EEG channels. For the analysis of frequencies higher than 30Hz spectral analyses of the EEG data were 449 performed using a sliding window multi-taper analysis. A window of 200ms length was 450 shifted over the data with a step size of 50ms with a spectral smoothing of 15 Hz. Single 451 trial z-normalization over the whole trial duration was performed (Grandchamp and 452 Delorme, 2011) and afterwards, a z-baseline was calculated with a window of 500ms 453 before cue onset (i.e. cue locked: -650ms to -150ms; pain locked: -2800ms to -2300ms for 454 low pain, -2900ms to -2400ms for medium pain and -2950ms to -2450ms for high pain). 455 performed at the electrode level were replicated in source space for display purposes and 480 interpolated to a 256x256x256 voxel space to identify potential neuronal generators. 481

482
Similar to a previous fMRI study (Fazeli and Buechel, 2018), our full model was comprised 483 of three experimental within-subject factors (see Fig. 2). The stimulus intensity factor 484 (INT) models the measured response with a simple linear function of the stimulus 485 intensity. The expectation (EXP) factor was defined (see Fig. 2; center column) linearly 486 from the intensity predicted by the cue. Again, conditions with a low pain cue were coded 487 with a -1, conditions with a medium pain cue with a 0 and conditions with a high pain cue 488 with a 1. The absolute prediction error factor (PE) resulted from the absolute difference 489 of the expectation and actual stimulus intensity (see Fig. 2; right column). 490

491
Behavioral pain ratings were averaged for all 3x3 cue-stimulus combinations over each 492 participant, resulting in a 35x9 matrix (subject x condition). We tested for main effects 493 across stimulus intensity, expectation, as well as prediction error using rmANOVA as Cluster permutation tests are specifically useful for explorative testing, as explained by 501 Maris & Oostenveld (2007). While prior hypotheses could have been formulated 502 regarding the spatial, temporal and spectral characteristics of brain responses associated 503 with the intensity of thermal stimulation and regions-of-interest could have been 504 described, variations in the present design could be related to temporal and spectral 505 differences compared to other studies which would be taken into account using non-506 parametric cluster permutation testing. 507 We wanted to explore positive and negative time-frequency patterns associated with our 508 variations of stimulus intensity, expectation and absolute prediction errors using a 509 repeated measures ANOVA. A statistical value corresponding to a p-value of .05 (F(2,28) 510 = 4.196) obtained from the repeated measures ANOVA F-statistics of the respective main 511 effect was used for clustering. Further, clustering was restricted in a way that only 512 samples were included in a cluster which had at least one significant neighbor in electrode 513 space, i.e. at least one neighboring channel also had to exceed the threshold for a sample 514 to be included in the cluster). 515 Subsequently, a cluster value was defined as the sum of all statistical values of included 516 samples. Monte Carlo sampling was used to generate 1000 random permutations of the 517 design matrix and statistical tests were repeated in time-frequency space with the 518 random design matrix. The probability of a cluster from the original design matrix (p-519 value) was calculated by the proportion of random design matrices producing a cluster 520 with a cluster value exceeding the original cluster. This test was applied two-sided for 521 negative and positive clusters, which were differentiated by the average slope of the 522 estimated factors. The cluster permutation test was done in a time frame from 0 to 1.6s 523 from pain onset in order to restrict the analysis to the pain duration in spite of the shift in 524 the time axis accounting for the ramp-up time of the thermode. 525