Reactivation of Single-Episode Pain Patterns in the Hippocampus and Decision Making

Experimental design

The experiments were designed to allow an investigation of the cognitive and neural mechanisms that support memory for aversive experiences, which is the focus of the current report. Secondarily, and separately, the experiment enables the investigation of the behavioral and neural correlates of pain modulation of immediate and very long-term recognition memory. For the latter question, a subset of participants returned 1 year later to assess whether the maintenance of recognition memory was modulated by pain and neural activity during the fMRI session; these results will be published separately (Wimmer and Buchel, 2015).

As an overview, the behavioral and fMRI experiments each started with a heat calibration phase. This was followed by the incidental learning phase (which was scanned in the fMRI experiment), where an abstract cue probabilistically associated with high or low heat was followed by the presentation of a trial-unique object in conjunction with high or low heat pain. A test phase followed that measured whether pain value memory could support rewarded choice of a low pain-associated object over a high pain-associated object (in the behavioral experiment) and single-item pain value memory (in the fMRI experiment).

The test phase was designed to be as sensitive as possible to behavioral signatures of pain value memory established via single episodes. Thus, we explicitly instruct participants to retrieve pain value associations. Such instruction differs from decisions about well-learned value associations, which can frame instructions in terms of preference. However, our use of many diverse stimuli prevents the use of such a general preference question: participants are bound to have idiosyncratic and widely varying preferences for the object pictures themselves. Thus, if a preference instruction had been used, these idiosyncratic preferences would be likely to dominate behavioral and neural responses at test phase re-exposure. Further, because of the requirement for novel experiences, a pre-rating phase to collect baseline object ratings was not possible.

The test phase in the fMRI experiment allowed for the investigation of the critical question of whether pain-related patterns of activity were reactivated on re-exposure to objects and whether individual differences in reactivation related to individual differences in value memory performance. Here, we used multivariate methods, which are powerful and highly sensitive tools for investigating neural representations in memory and of pain experience (Poldrack, 2011; Rissman and Wagner, 2012; Wager et al., 2013).

Heat calibration

Before the incidental learning phase, heat levels were calibrated for each participant to achieve the same subjective high and low aversive pain experience across participants. Thermal stimulation was delivered via an MRI-compatible 3 × 3 cm Peltier thermode (MSA; Somedic) applied to the inner left forearm. During the visual presentation of a white square, heat was applied for 10 s. Pain ratings were recorded with a 1-8 rating scale with 0.5 point increments, superimposed on a yellow-to-red gradient (as depicted in Fig. 1a). Participants moved an arrow cursor from the initial mid-point starting location using left and right key-presses and confirmed the rating by pressing the spacebar. A rating of 8 corresponded to the highest level of heat pain a participant could endure multiple times. If the level of pain was intolerable, participants moved the rating past the 8 end of the scale, at which point a 9 appeared on the screen. Participants rated the pain associated with a pseudo-random list of 10 different temperatures ranging from 39.5°C to 49.5°C. A linear interpolation algorithm then selected a low temperature estimated to yield a 2 rating and a high temperature estimated to yield a 7.5 rating.

To ensure that no damage to participants' skin was caused by the administered heat stimulation, the maximum temperature allowed in the experiment was 50.5°C. Further, as described in detail below, if participants at any point entered a 9 rating during the experiment, the high temperature was subsequently decreased by 0.8°C.

Procedure: incidental learning phase

In the incidental learning phase, participants experienced high or low heat pain while being exposed to trial-unique object pictures (Fig. 1a; common to both the behavioral and fMRI experiments). In the fMRI study, this phase was conducted inside the fMRI scanner.

Importantly, the encoding of the object pictures was incidental (not instructed) to more closely resemble the incidental nature of encoding in many real-world situations. Regarding the incidental object pictures, participants were given slides with the following instructions in text: “In the middle of the screen you will see object pictures during the experiment that you can just look at.” Later, they were instructed: “Attention test: If the shapes or the object pictures blink, please press the spacebar. The object pictures are there to keep your attention on the screen during the heat stimulus.” Color pictures of objects were drawn from a database of images compiled via Internet search (as used previously; Wimmer and Buchel, 2016); objects were largely composed of familiar nonfood household items set on white backgrounds.

Heat pain was probabilistically cued (70% predictive) to allow for some prediction of pain but also for surprise at pain onset, with a design adapted from Atlas et al. (2010) (see also Geuter et al., 2017; Fazeli and Buchel, 2018). Across four blocks, 33 included high heat trials and 33 included low heat trials were presented (of 35 total; Fig. 1a), including 10 low-to-high and 10 high-to-low mismatch trials. To allow for initial adjustment to the task, data from two initial low heat trials were excluded. In the behavioral choice study, these objects were omitted from the choice phase or in the fMRI study, these objects were presented first and then excluded from analysis. Given the low number of mismatch trials and the relatively low and inconsistent effect of cues on ratings (see Results), all analyses focused on administered heat regardless of cued heat to ensure reliability of imaging estimates.

To maintain attention on the screen during object presentation, participants were instructed to respond to occasional flickers in image brightness. The visual cue illumination flickered (decreased in illumination) once for 0.35 s in a random 50% of trials. Flicker timing was randomly distributed throughout the first 1.5 s of visual cue presentation. Similarly, in a separately determined random 50% of trials, the object picture flickered in illumination during heat stimulation. When either a visual cue or object flicker was detected, participants were instructed to press the down button.

To detail the timing of events in an incidental learning phase trial, first, a visual cue signaling likely high or low heat was presented for 2.5 s. Participants responded to a visual flicker if one occurred. After a 4 s interstimulus interval (ISI), the incidental object appeared. The incidental object was presented for a total duration of 10 s. Participants responded to a visual flicker if one occurred. Following the heat stimulation and after a 4 s ISI, a pain rating scale appeared. Participants used left and right buttons to move a selection arrow from the initial cursor position (randomized between 4.5 and 5.5) to their experienced pain level and pressed the down button twice to make their selection; responses were self-paced. After the participant entered the response, trials were followed by a variable 2 s mean (range: 0.5-6 s) intertrial interval (ITI).

To allow for a better match between the appearance of the object and the onset of noticeable heat, heat onset started 0.75 s before object appearance (for a similar method, see Forkmann et al., 2013). The thermode temperature increased from baseline (33°C) to the desired temperature at a rate of 5 degrees per second, which translates to ∼3.5 s to reach the range of the high heat temperature. After the 10 s object presentation period, the thermode temperature decreased at a similar rate. Thus, for low heat trials where the thermode did not need to reach a high value, the temperature during the 10 s presentation of the object was approximately constant at the desired value. For high heat trials, it took up to 2.5 s at the beginning of the 10 s period for the thermode to reach the peak. After each incidental learning phase block, the thermode was moved to a new location on the inner arm to avoid sensitization.

To maintain similar differences in subjective experience between the high and low heat conditions, temperatures were automatically adjusted throughout the task to maintain the targeted pain rating values. If the median of the previous 6 validly cued low heat trials fell below a rating of 1.5, the low temperature was increased by 0.2°C; if the median rating was >3, the low temperature was decreased by 0.2°C. For the high temperature, if the median rating fell to <7.5, the high temperature was increased by 0.2°C (if the temperature was <50.5°C). If a rating of 9 was given, indicating an intolerably high level of pain, the high temperature was decreased by 0.8°C. Such online adjustments of administered temperature are not commonly used in pain research that focuses on effects of expectation or placebo (e.g., Atlas et al., 2010), as in these cases administered temperature needs to be constant across the task. However, our focus here was on the subjective response to pain; thus, online adjustment allowed us to maintain very similar subjective responses to the majority of high and low heat stimuli.

Two pseudo-random orderings of incidental object pictures were used for counterbalancing object and heat associations. The assignment of abstract circles to high and low heat was also counterbalanced across participants. Further, after the first two blocks of the experiment, two new abstract circles were used as cues, with visual and verbal instruction about the new cues preceding the block. Visual cues were probabilistically associated with the level of heat, correctly predicting high or low heat in 70% of trials (Atlas et al., 2010). On invalid trials, the alternative heat level was administered. Additionally, 6 trials included a probe of cue-related pain expectancy: after 2.5 s of cue presentation, a question appeared below the cue asking participants whether they expected low or high heat to follow. These probes were used to ensure participants remained aware of the cue-pain associations. After the probe, trials continued as normal.

Procedure: behavioral choice test phase

In the behavioral study, a surprise choice test phase followed the incidental learning phase to examine value memory for the objects incidentally associated with high or low heat in the preceding phase. Participants were instructed to select the object, out of two alternatives, that was associated with lower heat pain in the preceding phase. One object had been associated with the administration of high heat (independent of the cue) and the alternative object that had been associated with low heat (independent of the cue). Participants were instructed that they could win €0.50 euro for each correct choice of the lower heat object on top of their payment for participation.

The choices sampled each of the 66 objects from the incidental learning phase without repetition, resulting in 33 choices. Choices were presented in a pseudo-random order. A given choice included either 2 objects that had been correctly cued to be of low and high heat or a choice between one validly cued object and one invalidly cued object. We found no influence of the invalid cue or whether pain was higher or lower than expected on choice accuracy (p values > 0.31), so we collapse across this factor in all analyses. Following these choices, an additional four trials presented choices between the abstract circle cues that had been predictive of high versus low heat pain.

To detail the timing of events on a choice trial, first, the choice options were presented serially in a random order (Fig. 1b). The first option was presented either on the left side or on the right side of the screen (determined at random) for 4 s, followed by a 1 s ISI. The second option was then presented in the alternate spatial location for 4 s, followed by a 1 s ISI. Then the first option returned to the screen, below the prompt “Lower heat? (€0.50 reward).” Participants could select the onscreen option by pressing the space key, or press the left or right key to alternate between the options. Alternation was allowed for an unlimited amount of time. After choice entry, a confidence rating followed, presenting the options: “Guess,” “Low,” “Medium,” and “High”. Participants responded using the 1-4 keys. A variable 3 s ITI followed.

Procedure: fMRI memory test phase

In the fMRI study, a surprise memory test followed the incidental learning session. While collecting fMRI data, we assessed memory for the level of pain experienced with the object and recognition memory strength (Fig. 1c). Participants saw each of the “old” objects from the incidental learning phase. The old objects were intermixed with 20 “new” objects for a total of 86 included trials.

The participants were given slides with text instructions, which included the following: “Try your best to indicate the heat strength that you remember being associated with them. It is likely that this is very difficult for you. Please just give your best guess or gut feeling. It is likely that you remember more than you think.” Before the start of this phase, they were reminded: “The heat question can seem difficult, but it's very important to the experiment, so try to do your best. Guessing is okay!” For the recognition memory strength responses, participants were given the following instructions: “You have already seen most of the pictures in the first part, but some are also 'new.' For the new pictures, it doesn't matter what you say about the heat rating and the question of how sure you are.”

In our test phase trials, the pain association response was collected first, before the control measure of recognition strength, a reverse in question order compared with common memory paradigms. This key feature was explicitly designed to allow us to best detect behavioral and neural evidence of pain value memory, and was motivated by multiple considerations. First, our design focused participants on any pain memory associations immediately on the re-presentation of an object to increase behavioral performance and to best temporally isolate neural reactivation of pain, which we expected to be triggered immediately on object re-presentation. Second, this order maximizes data collection by avoiding the loss of value memory responses for items rated as “new” as in common designs. Third, we prioritized value memory over the control recognition measure as our previous results found no link between reward value memory and recognition (Wimmer and Buchel, 2016). Fourth, our design facilitates generalization to decisions outside the laboratory: when making choices in the outside world, decisions about avoidance or approach can progress independently from recognition, and the ability to make such a determination quickly is likely to be adaptive.

To detail the timing of events on a memory phase trial, first, a single object was presented alone for 5 s. Next, after a 1 s ISI, an unmarked yellow-to-red heat scale with superimposed left- and right-pointing arrows was shown. Participants pressed the left or right buttons to indicate whether they thought that the object had been associated with low heat pain or high heat pain in the incidental learning phase. For objects that participants definitely considered to be “new,” participants were told that they could pick either the high or low heat response at random. If they were not sure whether an object was new, participants were instructed to try to recall the level of heat it may have been paired with. All test phase responses were self-paced. Next, a confidence rating screen appeared with four levels of response: “guess,” “somewhat certain,” “certain,” and “very certain.” For stimuli participants believed were definitely new and thus had no associated heat experience, participants were instructed to respond with a low confidence answer. After a variable ISI (mean: 4 s; range: 3-6.5 s), a 6 point memory recognition strength scale was presented (e.g., Schwarze et al., 2012). Participants indicated whether they thought the object was “new” (not previously seen) or “old” (seen during the learning task) with 6 levels of response: “certain new,” “somewhat certain new,” “guess new,” “guess old,” “somewhat certain old,” and “certain old.” Participants used the left and right buttons to move from the randomly initially highlighted “guess new” or “guess old” response option to their selected response and then pressed the down button twice to make their selection. A variable ITI with a mean of 4 s (range: 2-8 s) followed.

The order of the old pictures was pseudo-randomized from the incidental learning phase order, and the old and new pictures were pseudo-randomly intermixed. The duration and distribution of ITIs (or “null events”) were optimized for estimation of rapid event-related fMRI responses as calculated using Optseq software (http://surfer.nmr.mgh.harvard.edu/optseq/).

At the end of the experiment, participants completed a written questionnaire querying their knowledge of the task instructions and their expectations (if any) regarding the incidental object pictures. Task instructions and onscreen text were presented in German for all parts of the experiment; for the figures and methods, onscreen text has been translated into English.

Data acquisition

The experiment was presented using MATLAB (The Mathworks) and the Psychophysics Toolbox (Brainard, 1997). For the behavioral study and the pain calibration phase of the fMRI study, data were collected using a 15 inch Apple Macbook Pro laptop. Responses were made using left and right arrow keys and the space key. In the scanner for the fMRI study, the task was projected onto a mirror above the participant's eyes. Responses were made using a 4 button interface with a “diamond” arrangement of buttons. Skin conductance was recorded from the hypothenar of the left hand. The signal was amplified using a CED 2502 amplifier and digitized at 200 Hz using a CED micro1401 (both by Cambridge Electronic Design) and downsampled offline to 100 Hz.

Whole-brain imaging was conducted on a Trio 3 Tesla system equipped with a 32-channel head coil (Siemens). Functional images were collected using a gradient echo T2*-weighted EPI sequence with BOLD contrast (TR = 2460 ms, TE = 26 ms, flip angle = 80; GRAPPA factor of 2; 2 × 2 × 2 mm voxel size; 40 axial slices with a 1 mm gap). Slices were tilted ∼30° relative to the AC-PC line to improve signal-to-noise ratio in the orbitofrontal cortex (Deichmann et al., 2003). Head padding was used to minimize head motion; no participant's motion exceeded 3 mm in any direction from one volume acquisition to the next. For each functional scanning run, five discarded volumes were collected before the first trial to allow for magnetic field equilibration.

During the incidental learning phase, four functional runs of an average of 190 TRs (7 min and 48 s) were collected, each including 17 trials. During the memory test phase, four functional runs of an average of 196 TRs (8 min and 2 s) were collected, each including 22 trials. If a structural scan had been collected for the participant at the center within the past 6 months, the previous structural scan was used. If not, structural images were collected using a high-resolution T1-weighted MPRAGE pulse sequence (1 × 1 × 1 mm voxel size) between the incidental learning phase and the immediate memory test phase (12 participants). We found no relationship between the consequently varying time delay between the end of the incidental learning phase and the start of the test phase on value memory performance, recognition memory performance, or fMRI pain reactivation measures (from n = 28 participants with timing and structural scan origin information).

All voxel locations are reported in MNI coordinates, and results are displayed overlaid on the average of all participants' normalized high-resolution structural images using the xjView toolbox (http://www.alivelearn.net/xjview) or AFNI (Cox, 1996).

Behavioral analysis

Our primary behavioral question was whether memory-based decisions were influenced by the pain that had been experienced with objects in the preceding incidental learning phase. In the behavioral experiment, choice trials were excluded if the administered heat for the high heat stimulus did not exceed that for the low heat stimulus (in rare cases when the thermode failed to increase temperature; on average <1 trial per participant).

In both experiments, we conducted simple a priori comparisons of behavioral performance to chance (50%) using t tests, with a significance threshold of p < 0.05 (two-tailed). We also examined the influence of cue expectation on pain ratings using a paired t test. In the fMRI experiment, we further verified in initial comparisons that “old” objects (whether paired with high or low pain) were recognized at a higher rate than “new” objects.

To further investigate value memory, multilevel regression models were implemented in R using lme (from the nlme package) for linear regression and glmmTMB (from the glmmTMB package) for logistic regression. All predictors and interactions were included as random effects, following the “maximal” approach (Barr et al., 2013). In all regressions, participant was entered as a random effect along with all other variables of interest. Correlations between random effects were included when convergence was achievable with this structure. All reported p values are two-tailed. In a control model, we verified that the presence versus absence of a visual “flicker” during object presentation was not related to value memory or recognition memory strength.

We additionally tested whether nonsignificant results were weaker than a medium effect size using the two-one-sided t test (TOST) procedure (Schuirmann, 1987; Lakens, 2017) and the TOSTER library in R (Lakens, 2017). In the behavioral experiment (n = 21), we used bounds of Cohen's d = 0.64, where power to detect such a medium effect is estimated to be 80%. In the larger fMRI sample (n = 29), we used bounds of Cohen's d = 0.55 to achieve the same estimated power.

fMRI preprocessing

Preprocessing and data analysis were performed using Statistical Parametric Mapping software (SPM12; Wellcome Department of Imaging Neuroscience, Institute of Neurology). Before preprocessing, slices with artifacts were identified as having mean image intensity ≥5% above the across-run slice mean. Individual slices with artifacts were replaced with the mean of the two surrounding time points using a script adapted from the ArtRepair toolbox (Mazaika et al., 2009). Images were then slice-timing corrected, realigned to correct for participant motion, and then spatially normalized to the MNI coordinate space by estimating a warping to template space from each participant's anatomical image and applying the resulting transformation to the EPIs. Images were filtered with a 128 s high-pass filter and resampled to 2 mm cubic voxels. Images were then smoothed with a 6 mm FWHM Gaussian kernel for univariate and connectivity analyses.

fMRI univariate analyses

fMRI model regressors were convolved with the canonical HRF and entered into a GLM of each participant's fMRI data. The six scan-to-scan motion parameters produced during realignment were included as additional regressors in the GLM to account for residual effects of participant movement. All regressions were conducted with automatic orthogonalization in SPM turned off.

Our primary univariate analysis was a “localizer” analysis to identify main effects of pain during the incidental learning phase. The GLM included regressors for the cue period (2.5 s duration), the initial pain onset period (2 s), the full pain and object presentation period (10 s), and the pain rating period (with a variable duration based on response time). The cue period regressor was accompanied by a parametric modulator contrasting high versus low expected pain. The pain onset period regressor was accompanied by two parametric modulators: the mismatch between cue and pain as well as the unsigned (absolute value) mismatch between cue and pain (these regressors were not correlated; r = 0.007). The full pain period regressor was accompanied by a parametric modulator representing the pain rating given on that trial. The regions identified as correlating with pain during the 10 s pain period were the same with or without the inclusion of the 2 s pain onset regressor.

Then we conducted several control univariate analyses. First, we examined learning phase activity correlated with later successful pain value memory and recognition memory strength. This model was based on the GLM above, but instead of the pain rating parametric modulator, we included parametric modulators for subsequent correct value memory and subsequent recognition memory strength. Separate parametric regressors were used for high and low pain-associated objects to allow for baseline differences (yielding four parametric regressors in total); results were then combined at the second level.

The remaining control univariate analyses examined activity in the test phase. In these models, a 5 s regressor modeled activity during the object re-presentation period. Additional regressors modeled the pain memory response period, the pain memory confidence response period, and the memory response period; the durations for all these periods matched the participant's response time. First, we looked for univariate correlates of pain reactivation to confirm that any multivariate results were not primarily driven by univariate activity. Thus, the object re-presentation regressor was accompanied by a parametric modulator representing the level of heat pain experienced with objects in the preceding learning phase. Second, a control test phase univariate analysis examined correlates of pain value memory success and recognition memory strength. Here, the object re-presentation regressor was accompanied by parametric regressors representing value memory success and recognition memory strength; as in the learning phase, separate regressors were used for high and low pain-associated objects and were combined at the second level.

Multivariate fMRI analyses

To test our primary fMRI prediction that patterns of BOLD activity associated with negative emotional experience were reactivated at retrieval, we used multivariate classification analyses. These analyses used the nonsmoothed fMRI data. In the incidental learning phase and the memory test phase, we estimated mass-univariate GLMs where each trial was modeled with a separate regressor. For the incidental learning phase, each regressor modeled the onset of an object and continued through the 10 s duration of the heat stimulus. For the memory test phase, each regressor began at the onset of the object and continued for the 5 s duration of object presentation (before any responses). Models included the six nuisance motion regressors (translations and rotations).

Multivariate analyses were conducted using The Decoding Toolbox (Hebart et al., 2014). Classification used an L2-norm learning support vector machine (LIBSVM; Chang and Lin, 2011) with a fixed cost of c = 1. The classifier was trained on the full incidental learning phase balanced via bootstrapping. The trained classifier was then tested on the full memory test phase data. For the primary across-phase classification analysis, no cross-validation is necessary for training because no inferences are drawn and no results are reported from the incidental learning phase data. Memory test phase classification is reported as area under the curve (AUC), which uses graded decision values and better accounts for biases in classification that may arise because of the different processes engaged by the incidental learning and memory test phases. Supplemental ROI analyses examined training and testing within the learning phase or memory test phase using cross-validation. Using cross-validation, we computed the strength of discriminability in the localizer phase in our ROIs.

Additionally, we conducted a searchlight analysis for further localization. We used a 4 voxel radius spherical searchlight (∼208 voxels). Training of the classifier on the incidental learning phase and testing on the memory test phase were conducted as described above for the ROI multivoxel pattern analysis. Individual subject classification accuracy maps were smoothed with a 6 mm FWHM kernel before group-level analysis. We also performed covariate analyses to determine whether behavioral performance was correlated with classification accuracy.

It has been shown that it is not valid to conduct statistical inference specifically on cross-validated classification accuracy measures of information using t tests (Allefeld et al., 2016). In part, as informational measures cannot be below zero, assumptions underlying the t test are violated for cross-validation within the same dataset. Our classifier training and testing were conducted on separate datasets (“cross-classification” between the incidental learning and the memory test phase), which does allow for potential “true” below-zero values, a case not addressed by Allefeld et al. (2016). Further, we found that cross-classification AUC values in all our ROIs followed a normal distribution (Anderson-Darling goodness-of-fit hypothesis test). While the above concern may still apply to inferences made about the main effects of pain during the incidental learning phase, our primary hypothesis rests on the cross-classification of pain-related patterns from the memory test phase.

Connectivity analyses

We additionally conducted psychophysiological interaction (PPI) analyses to examine differences in functional connectivity for successful versus unsuccessful value memory retrieval. These analyses used a hippocampal ROI as the seed region (defined in Results). In the incidental learning phase, we estimated a PPI contrasting correct versus incorrect later value memory retrieval, modeling the 10 s duration of the object and pain period. In the memory test phase, we estimated a similar PPI analysis, contrasting correct versus incorrect value memory retrieval, modeling the 5 s duration of the object presentation period. At the second level, we performed correlation analyses to determine whether behavioral performance was related to differences in connectivity for correct versus incorrect encoding or retrieval of value memory associations.

Statistical correction and ROIs

For both univariate and searchlight results, linear contrasts of univariate SPMs were taken to a group-level (random-effects) analysis. We report results corrected for family-wise error (FWE) due to multiple comparisons (Friston et al., 1994). We conduct this correction at the peak level within small-volume ROIs for which we had an a priori hypothesis or at the whole-brain cluster level (in each case using a cluster-forming threshold of p < 0.005 uncorrected, except for the pain rating correlation, where we used p < 0.00001 to yield more interpretable clusters).

We focused on two a priori ROIs motivated by two separate hypotheses. Given the anterior insula's role in processing the affective qualities of pain (Kurth et al., 2010; Wiech et al., 2014), we predicted that the insula may relate to the modulation of memory by pain. For this pain hypothesis-motivated anterior insula ROI, we first created a bilateral anterior insula mask (Brooks et al., 2002; Wiech et al., 2014), covering the insular cortex anterior to y = 9, as well as up to 4 mm lateral or superior to the insular cortex to account for signal blurring and anatomical variability. This mask was further restricted by the main effect of pain rating taken from the incidental learning phase localizer GLM defined above, thresholded at p < 0.0001 uncorrected (https://neurovault.org/images/390703/). We also defined a broader pain-related mask based on the localizer GLM thresholded at p < 0.0001 uncorrected, excluding the cerebellum. Separately, we focused on the hippocampus because of its role in episodic and relational memory (Eichenbaum and Cohen, 2001; Davachi, 2006). We also conducted follow-up analyses in the anterior hippocampus, given its role in negative emotion-related memory and generalization (Fanselow and Dong, 2010; Poppenk et al., 2013). The bilateral hippocampus ROI was derived from the Harvard-Oxford atlas at a threshold of 50%. We focused on a restricted mask of the hippocampus to limit the size of the ROI for multivariate analyses. We confirmed that there was no overlap between the hippocampus and pain-related masks. We defined the anterior hippocampus as the mask region anterior to Y = −21, approximating the position of the uncal apex (Poppenk et al., 2013). While somatic processing of thermal pain does not primarily involve the amygdala, as a control we also examined the amygdala, defined from the Harvard-Oxford atlas at a threshold of 50%.

Correlations between classification accuracy and behavioral performance were conducted using Pearson' correlation. Statistical comparison of the difference between correlations was computed using Steiger's test for differences in dependent correlations.

Data availability

Behavioral data are available on the Open Science Framework (https://osf.io/gr9xd/). Whole-brain fMRI results are available on NeuroVault (https://neurovault.org/collections/6126/).