Identifying inter-individual differences in pain threshold using brain connectome: a test-retest reproducible study

Abstract

Individuals are unique in terms of brain and behavior. Some are very sensitive to pain, while others have a high tolerance. However, how inter-individual intrinsic differences in the brain are related to pain is unknown. Here, we performed longitudinal test-retest analyses to investigate pain threshold variability among individuals using a resting-state fMRI brain connectome. Twenty-four healthy subjects who received four MRI sessions separated by at least 7 days were included in the data analysis. Subjects’ pain thresholds were measured using two modalities of experimental pain (heat and pressure) on two different locations (heat pain: leg and arm; pressure pain: leg and thumbnail). Behavioral results showed strong inter-individual variability and strong within-individual stability in pain threshold. Resting state fMRI data analyses showed that functional connectivity profiles can accurately identify subjects across four sessions, indicating that an individual’s connectivity profile may be intrinsic and unique. By using multivariate pattern analyses, we found that connectivity profiles could be used to predict an individual’s pain threshold at both within-session and between-session levels, with the most predictive contribution from medial-frontal and frontal-parietal networks. These results demonstrate the potential of using a resting-state fMRI brain connectome to build a ‘neural trait’ for characterizing an individual’s pain-related behavior, and such a ‘neural trait’ may eventually be used to personalize clinical assessments.

1. Introduction

Despite the universality of acute pain and the high prevalence of chronic pain, characterizing the mechanisms of pain perception and modulation is still a fundamental challenge for pain researchers. Neuroimaging approaches, such as electroencephalography (EEG) and magnetic resonance imaging (MRI), provide opportunities for studying central processing of pain non-invasively and in vivo (Davis, 2011; Moayedi et al., 2018; Tracey and Mantyh, 2007). In the past decades, a large body of functional and structural neuroimaging studies have identified neural responses and signatures for both acute pain (Brodersen et al., 2012; Tu et al., 2016b; Wager et al., 2013; Woo et al., 2015; Zhang et al., 2012) and chronic pain conditions (Baliki et al., 2008; Bushnell et al., 2013; López-Solà et al., 2017; Simons et al., 2014; Ung et al., 2014; Yu et al., 2014)

To investigate pain-related brain responses, previous neuroimaging studies have normally collapsed data from many subjects and built a common map between neural activity and pain perception across subjects. However, pain sensitivity varies significantly across individuals (Mogil, 1999; Nielsen et al., 2005). Understanding and identifying the neural pattern of inter-subject variability are of high significance in both fundamental and clinical pain studies. Several studies have explored brain patterns associated with individuals’ pain characteristics using functional and structural MRI (Cheng et al., 2015; Coghill et al., 2003; Emerson et al., 2014; Wager et al., 2013) and EEG (Hu and Iannetti, 2019; Huang et al., 2013; Schulz et al., 2012, 2011). However, no reliable neural index that accounts for variability in pain perception across individuals has been established since previous findings were hindered by the use of unimodal pain measures (e.g., heat, laser stimuli) and lack of test-retest reliability.

Since pain is a complex and subjective experience involving sensory-discriminative, affective-motivational and cognitive-evaluation dimensions (Tracey, 2011; Wiech et al., 2008), it is natural to expect that an individual’s pain characteristics may be encoded in large-scale brain networks/connectome instead of single regions. Recent studies have demonstrated that the resting-state brain connectome is unique to each individual (Finn et al., 2015) and can robustly predict psychological traits (Beaty et al., 2018; Feng et al., 2018; Rosenberg et al., 2016). Indeed, although pain varies across individuals, pain characteristics (e.g., pain threshold and pain tolerance) have been reported to be relatively stable for an individual (Geber et al., 2011; Koenig et al., 2014). Therefore, we hypothesize that a reliable neural index of an individual’s pain characteristics could be identified from their resting-state brain connectome.

In the present study, we explored the within-subject stability and between-subject variability of the resting-state brain connectome and pain thresholds as well as their association. Twenty-four healthy subjects participated in the study and received MRI sessions separated by at least 7 days for about one month. Subjects’ pain thresholds were measured using two modalities of pain (heat and pressure) on two different locations (heat pain: leg and arm; pressure pain: leg and thumbnail). We hypothesize that 1) individuals may be identifiable by their brain connectome; 2) pain threshold may be predicted using the brain connectome; and 3) patterns encoding pain threshold are reproducible across the four MRI sessions.

2. Methods and Materials

2.1. Participants

Twenty-four participants (8 males; age = 25.2 ± 0.77) were included in the present study and were asked to maintain their usual daily activities for the duration of study involvement. The study was approved by the Partners Human Research Committee at Massachusetts General Hospital. All participants had normal or corrected-to-normal vision and gave written informed consent prior to participating in the study.

2.2. Pain Measurements

For each of the four visits, two pain modalities were assessed using quantitative sensory testing. Pain threshold assessments at two locations (leg and arm/thumbnail) were conducted 3 times, with the heat thermode and pressure algometer repositioned between each threshold assessment. We averaged pain thresholds across 3 separate tests to obtain a stable measurement. We chose 2 pain modalities because heat-evoked pain is predominantly mediated by small, nonmyelinated peripheral nociceptive nerve fibers (C-fibers), whereas pressure-evoked pain is predominantly mediated by small, myelinated peripheral nociceptive nerve fibers (A-delta fibers) (Angst et al., 2009). We measured pain threshold at both local and distal locations in order to compare within-modality/between-modality pain threshold scores and neural pain trait.

Contact heat stimuli were delivered using a PATHWAY CHEPS (Contact Heat-Evoked Potential Stimulator, Medoc Advanced Medical Systems), with pain thresholds measured on the medial side of the right knee and left volar forearm. An ascending method, with a rate of increase of 0.5 °C/s from 32 °C was applied. A study staff member held the thermode lightly on the skin. Participants were required to press a button to indicate when the heat stimulus first became painful, thereby indicating the heat pain threshold. Pressure pain thresholds were assessed using an algometer applied at the medial side of the right knee and left thumbnail. Pressure was gradually increased at a constant rate of 1 kg/s. The participant was instructed to say “stop” to indicate when the sensation first became painful.

In summary, we collected two ‘modalities’ (heat pain and pressure pain) and four ‘measures’ (heat pain on leg, heat pain on arm, pressure pain on leg, and pressure pain on thumbnail) of the 24 participants during each of the four visits. The STAI-state was administered before pain measurements to provide a measure of the participant’s current anxiety level.

2.3. MRI Acquisition

All functional MRI data were acquired using a 32-channel radio-frequency head coil in a 3T Siemens scanner at the Massachusetts General Hospital Martinos Center for Biomedical Imaging. During the resting-state fMRI, subjects were asked to keep their eyes open and to blink normally while looking at a darkened screen for approximately 8 minutes. A whole-brain gradient-echo echo-planar-imaging sequence was used for functional scanning with a repetition time (TR) of 3000 ms (30 ms echo time, 44 3.0 mm-thick slices, 2.6 × 2.6 mm in-plane resolution), and a total of 160 volumes were collected. A high-resolution, Tl-weighted structural image (1 mm³ isotropic voxel MPRAGE) was acquired after functional imaging.

2.4. Data Analysis

2.4.1. Behavioral Data Analysis

Two-way repeated measures analysis of variance (ANOVA), with gender (male and female) and visit (visit 1 to 4) as factors, was used to assess the difference in pain threshold for each subject across the four visits. This analysis was performed separately for four different measures of pain threshold. We also calculated correlations of pain threshold measures between two modalities for each visit. To rule out the potential effect of anxiety on pain threshold, we also calculated the correlations between STAI-state and pain threshold measures for each visit.

2.4.2. fMRI Data Preprocessing and Head Motion

The fMRI data were preprocessed using CONN toolbox v17C (http://www.nitrc.org/proiects/conn). The images were preprocessed with slice-timing, realigned, coregistered to subjects’ respective structural images, normalized, and smoothed with a 6 mm full width at half maximum (FWHM) kernel. Then, segmentation of gray matter, white matter, and cerebrospinal fluid (CSF) areas to remove temporal confounding factors was conducted. Band-pass filtering was performed with a frequency window of 0.01 to 0.1 Hz. To eliminate correlations caused by head motion and artifacts, we identified outlier time points in the motion parameters and global signal intensity using ART (https://www.nitrc.org/projects/artifact_detect). We treated images as outliers if the composite movement from a preceding image exceeded 0.5 mm or if the global mean intensity was greater than 3 standard deviations from the mean image intensity. Outliers were included as regressors in the first level general linear model along with motion parameters.

2.4.3. Longitudinal and Across-measures Test-retest Analyses

To identify the resting-state brain connectome that encoded inter-individual variability in pain threshold and test the replicability of the identified ‘neural trait’, we performed longitudinal and across-measure test-retest analyses.

For longitudinal test-retest analyses (Fig. 1A, 1), we used resting-state fMRI data collected from one of the four sessions for 24 subjects as the ‘database’ to build a model, and we tested the performance of the model on the data from the other three ‘target’ sessions. For example, we built a model from data collected from the first visit (denoted as ‘Rest 1’) and tested its performance on the other three visits (denoted as ‘Rest 2’, ‘Rest 3’ and ‘Rest 4’). It should be noted that the test-retest analyses were bi-directional. For instance, we built a model from ‘Rest 1’ and tested it on ‘Rest 2,’ and we also built a model from ‘Rest 2’ and tested it on ‘Rest 1.’ By assigning sessions to both ‘target’ and ‘database’ roles, we investigated twelve pairs of scan sessions for each subject.

Fig. 1. The framework of the study.

A. Study design. We collected resting-state fMRI data and four measures of pain threshold (heat pain on leg/arm, pressure pain on leg/thumbnail) from each subject at four different sessions over one month, with each session separated by at least 7 days. We performed longitudinal test-retest using the data from one session as database and a second session acquired in another week as the target set, and all possible combinations (bi-directional, 12 in total) of sessions are indicated by the arrows connecting session pairs. B. Identification procedure. Given a connectivity matrix from the target set, we computed the correlations between this matrix and all the connectivity matrices from the database. The predicted identity is the one with the highest correlation coefficient. C. Predict individual pain threshold scores. We built multivariate linear regression models with pain threshold as dependent variables and multivariate connectivity matrices as independent variables. The models were decoded using support vector regression (SVR) based on cross validation (within-session) and external validation (between-sessions). CV: cross validation; EV: external validation. D. Node and network definitions. We used a 268-node functional atlas defined in (Shen et al., 2013). Nodes were grouped into eight networks.

Across-measure test-retest analyses (Fig. 1A, 2) of pain thresholds were studied at three levels: within-measure (e.g., trained the model on ‘heat pain on leg’ and tested the model on the same measure), within-modality (e.g., trained the model on ‘heat pain on leg’ and tested the model on ‘heat pain on arm’), and between-modality (e.g., trained the model on ‘heat pain’ and tested the model on ‘pressure pain’).

Fig. 2. Pain threshold scores.

A. Boxplots of individuals’ pain threshold scores in different sessions. Each circle represents the score of an individual’s pain threshold (one of the four measures) in a session. The blue box represents the 25th and 75th percentiles, and the red line represents the median of pain threshold scores (one of the four measures) across subjects in a session. B. Correlation maps for different measures of pain threshold. The blue and white squares highlight the correlations between scores of heat pain on leg and heat pain on arm, and between scores of pressure pain on leg and pressure pain on nail, respectively.

2.4.4. Functional Parcellation and Network Definition

To build a resting-state brain connectome for each subject, we parcellated the brain using a 268-node functional atlas that was defined (Shen et al., 2013), tested (Finn et al., 2015; Rosenberg et al., 2016), and used in a recent protocol (Shen et al., 2017). Those 268 nodes were further categorized into one of eight networks (medial frontal, frontoparietal, default mode, subcortical-cerebellum, motor, visual 1, visual 2, and visual association; Fig 1D).

The Pearson correlation coefficients between the time courses of each possible pair of nodes were calculated and normalized to z-scores using the Fisher transformation, resulting in a 268×268 symmetrical connectivity matrix, in which each element represents a connection strength, or edge, between two nodes. This was done for each subject for each session separately, such that each subject had a total of four matrices reflecting connectivity patterns during each of the different scan sessions.

2.4.5. Identifying Individuals Using the Brain Connectome

Using an approach similar to that of a previous study, we tested the hypothesis that individual variability is encoded in the functional brain connectome and explored the robustness and reliability of identifying individuals using connectivity matrices (Finn et al., 2015).

Identification was performed across 12 pairs of sessions consisting of one ‘target’ and one ‘database’ session. Fig. 1b illustrates the identification procedure. In the first step, we created a database that consisted of all (N=24) subjects’ connectivity matrices from one session. Second, the identity of the target connectivity matrix from a different session was determined by calculating the similarity between the current target matrix and all other matrices (N=24) in the database, and the identity was defined as that with the maximal similarity score. Similarity was defined as the Pearson correlation between the edge values taken from the target matrix and each of the database matrices. This procedure ran 24 times iteratively to give identities for every subject in the target session. Once all subjects in the target session were given an identity, the identification rate was measured as the percentage of subjects whose identity was correctly identified out of the total number of subjects.

We also tested the identification rate using each of eight functional networks. A sub-matrix corresponding to a single network was used, and the same procedure as whole-brain connectome identification was performed to investigate the contribution of the individual network.

The statistical significance of identification accuracy was assessed by nonparametric permutation testing. In each iteration, we randomly selected one session from all sessions of all subjects to serve as the database and another session to serve as the target set. Then, subject identity in the database set was permuted, and each subject in the target set was compared with an ‘incorrect’ identity from the database. This procedure was repeated 1000 times to build the distribution of identification rate from randomly labeled subjects.

2.4.6. Predicting Individual’s Pain Threshold Scores

To determine whether individual differences in functional connectivity are relevant to individual differences in pain perception, we used brain profiles to predict subjects’ pain threshold scores both within and between sessions. The prediction analyses (Fig. 1C) included four major steps:

1. Feature selection. Generally, we used whole-brain connectivity matrices (see 2.4.4 for details) as multivariate features in the prediction analyses. Based on the results of identification analyses, we also tested the contributions of five single networks (medial frontal, frontoparietal, default mode, subcortical-cerebellum, and motor) by selecting sub-matrices as features in the prediction. These five networks have shown strong contributions for encoding interindividual variability in previous studies (Beaty et al., 2018; Feng et al., 2018; Finn et al., 2015) and our study (see Results for details), and more importantly, these networks support sensory-discriminative, affective-motivational, and cognitive-evaluation dimensions of pain (Tracey, 2011; Wiech et al., 2008). We did not include three visual networks since they were not substantial or useful in identifying individuals, and they were also less relevant to pain perception.

2. Model building. We built multivariate linear regression (MVLR) models with pain threshold as a dependent variable and multivariate functional connectivity as an independent variable. The models were built separately for different measures in each session. Such models have been widely used in fMRI studies to identify brain patterns related to pain-related behaviors (Lindquist et al., 2017; Tu et al., 2018, 2016b, 2016a; Wager et al., 2011; Zhang et al., 2019).

3. Model training and testing. We trained and tested the MVLR models using support vector regression (SVR, implemented by LIBSVM) based on cross validation and external validation. For cross validation, the models were trained within each session for different measures. Five-fold cross validation was used to ensure separation between training and testing samples. Specifically, we partitioned all subjects into five groups and used four groups for training and one group for testing. This procedure was repeated 5 times to make sure that each subject was used as the test sample once and the model did not include information from the test samples. For external validation, the models were trained from the data in one session and tested on the other three sessions. For example, we trained a model to predict individuals’ heat pain thresholds on the leg from the first session and applied the model to the data from the other three sessions to predict those pain thresholds.

4. Prediction. The predicted pain thresholds were calculated by taking the dot product of the prediction weights (obtained from training) and corresponding edge values from subjects in test samples. We assessed the predictive power of each model by correlating predicted and real pain thresholds across all subjects. This index represents how much variation of pain thresholds was explained by the predictive model (Lindquist et al., 2016; Tu et al., 2019b; Wager et al., 2013).

Given that individuals’ pain measurement scores were strongly correlated within pain modalities but not significantly related between modalities (See 3.1 for details), we summarized the patterns of functional connectivity that encoded individuals’ heat pain and pressure pain thresholds separately. Since we trained the model on one session and tested it on the other three sessions and this procedure was repeated four times to ensure that data from each session was used once as training data and three times as data to be tested, the models for each pain measure (e.g., heat pain on leg) were slightly different (e.g., edges with significant contributions for prediction). To identify edges with both significant and consistent contributions to predicting pain thresholds, we only retained edges showing significant contributions to prediction (by finding the overlap) across the eight models per modality (e.g., two measures [heat pain on leg/arm] × four sessions).

The statistical significance of the predictive power of each model was assessed by nonparametric permutation testing. In each iteration, we randomly permuted the labels of the data (pain threshold scores) prior to training, and we performed all procedures of prediction analyses on the permuted dataset. The permutation ran 1000 times to build the distribution of predictive power from randomly labeled subjects.

2.4.7. Additional Head Motion Analyses

Since head motion is a known confounder of connectivity analyses (Power et al., 2014), we performed several additional analyses to address this issue: First, instantaneous head motion was expressed by a scalar quantity, namely framewise displacement (FD) (Power et al., 2014). We used mean FD as the summarized head motion value for each subject. Second, we compared head motion values across the four sessions and correlated them with each individual’s pain threshold (four measures). Third, we included head motion values as a feature in the prediction model and compared the performance with and without head motion.

3. Results

3.1. Pain Threshold Scores

Two-way repeated measures ANOVA revealed no significant effect of visit, gender, and their interaction on pain threshold (Table 1), indicating that both females and males had stable pain threshold scores for all measures across the four sessions (Fig. 2A), and pain threshold did not differ between genders. The box plots in Fig. 2A show individual’s pain threshold scores and their median, 25th, and 75th percentiles values in each session, demonstrating that pain threshold scores varied largely across subjects. We also provide pain threshold data for each individual in Supplementary Figs. 1–4.

Table 1.

Two-way repeated measured ANOVA assessed pain threshold

	df	F	P
Heat Pain on Leg
Visit	3	2.21	0.10
Gender	1	3.77	0.09
Visit × Gender	3	0.06	0.98
Heat Pain on Arm
Visit	3	2.86	0.06
Gender	1	1.06	0.34
Visit × Gender	3	0.14	0.93
Pressure Pain on Leg
Visit	3	0.68	0.57
Gender	1	0.20	0.67
Visit × Gender	3	0.80	0.51
Pressure Pain on Nail
Visit	3	0.95	0.44
Gender	1	0.24	0.65
Visit × Gender	3	1.32	0.30

We calculated the correlations between different measures of pain threshold in each session. Fig. 2B shows the correlation maps. We found that within-modality measures (heat pain on leg vs. heat pain on arm; pressure pain on leg vs. pressure pain on thumbnail) were strongly correlated (p < 0.001 for all correlations; Fig. 2B, blue and white squares highlighted), while between-modality measures were not strongly correlated, indicating that heat pain and pressure pain may have different mechanisms.

We found that pain threshold scores were not correlated with state anxiety scores (STAI-state, p > 0.05 for all correlations), indicating that anxiety was not a confounding factor in the present study. It is worth mentioning that the higher the pain threshold for a subject, the higher the temperature or pressure recorded when the subject’s sensation first became painful.

3.2. Identifying Individual Subjects Using the Brain Connectome

3.2.1. Whole-Brain Identification

We used the whole-brain connectivity matrix to identify individual subjects. The identification rates are summarized in Fig. 3. The success rate was 71% (17/24) for both Rest 1-Rest 2 and the reverse Rest 2-Rest 1 database-target pairs. The success rate ranged from 71% to 88% for the other database-target pairs (Fig. 3B). The overall rate of whole-brain identification across all possible pairs was 78.5%. We did not see a trend between the interval time of database-target pair and identification rate (one week: 78%; two weeks: 80%; three weeks: 77%).

Fig. 3. Identification rates across session pairs and networks.

A. Identification rate based on all pairs of database and target set. Grey or black indicates which session was used as the database with the other session serving as the target. Bars show success rates based on whole-brain connectivity matrix (All), as well as one of eight individual networks (1–8). B. Summarized rates of whole-brain connectivity identification for all 12 pairs of database and target sessions.

Nonparametric permutation testing showed that the highest rate for identification using randomly labeled databases was 21%. Therefore, the p-value associated with obtaining an identification rate higher than 71% was < 0.001.

3.2.2. Single-Network Identification

We tested the identification rate using each of the eight functional networks to explore if certain brain networks contribute more to discriminating between individual subjects. Similar to the findings of Finn et al., we found that the medial frontal and frontal parietal networks (including the frontal, parietal, and temporal lobes; Fig. 1A, networks 1 and 2) had higher and more consistent success rates of identification (Finn et al., 2015). In particular, the medial frontal network showed the highest rates in all possible database-target pairs.

3.3. Predicting Individuals’ Pain Threshold Scores

Given that 1) individuals could be identified by functional connectivity and 2) there was large variability in pain thresholds across subjects but stable scores across sessions, we then tested whether connectivity could be used to predict subjects’ pain threshold scores.

3.3.1. Within-Session Prediction

Fig. 4 shows the edges of connectivity with significant contribution (Bonferroni-corrected for multiple comparisons across edges) to the prediction of heat pain thresholds, obtained from two measures (leg and arm) from four sessions. We mainly observed that connectivity between the cerebellum and prefrontal, temporal, and motor lobes was positively associated with heat pain threshold, while connectivity between the temporal and limbic lobes and between the parietal and subcortical lobes was negatively associated with heat pain threshold.

Fig. 4. Predicting individual’s heat pain threshold using functional connectivity.

Upper panel: edges of connectivity with significant contribution (Bonferroni-corrected for multiple comparisons across edges) to the prediction of heat pain threshold, obtained from two measures (leg and arm) from four sessions. Red/blue lines denote positive/negative relationships between strength of edges and pain threshold scores. Lower panel: performance of predictions using whole-brain or single-network connectivity in four sessions. Red dashed lines represent the 95% confidence interval obtained from nonparametric permutation testing.

The lower panel of Fig. 4 shows the performance (indexed by the correlation between predicted and real scores of pain threshold) of predicting pain threshold scores using whole-brain or single-network connectivity in different sessions. We found that the predictions using whole-brain connectivity, medial frontal network (network 1), frontoparietal network (network 2), and subcortical-cerebellum network (network 4) could significantly (higher than the 95% confidence interval obtained from nonparametric permutation testing; red dash lines in Fig 4, low panel) predict individuals’ heat pain threshold scores on both leg and arm in all four sessions.

Fig. 5 shows the edges of connectivity with significant predictive power for pressure pain threshold (Bonferroni-corrected for multiple comparisons across edges), obtained from two measures (leg and thumbnail) and four sessions. We mainly observed that connectivity from the frontal and parietal lobes was positively associated with pressure pain threshold, while connectivity from the temporal lobes was negatively associated with pressure pain threshold.

Fig. 5. Predicting individual’s pressure pain threshold using functional connectivity.

Upper panel: edges of connectivity with significantly contribution (Bonferroni-corrected for multiple comparisons across edges) to the prediction of pressure pain threshold, obtained from two measures (leg and thumbnail) from four sessions. Red/blue lines denote positive/negative relationships between strength of edges and pain threshold scores. Lower panel: performance of predictions using whole-brain or single-network connectivity in four sessions. Red dashed lines represent the 95% confidence interval obtained from nonparametric permutation testing.

The lower panel of Fig. 5 shows the performance of predicting pain threshold scores using whole-brain or single-network connectivity in different sessions. We found that the predictions using whole-brain connectivity, medial frontal network (network 1), and frontoparietal network (network 2) could significantly predict individuals’ pressure pain threshold scores on both the leg and thumbnail in all four sessions.

3.3.2. Between-Session Prediction

We next tested the model’s performance in predicting individuals’ pain thresholds using whole-brain connectivity between different sessions. The models were trained from the data in one session and tested on the other three sessions. Fig. 6A shows the summarized values of performance for within-session and between-session predictions. The performance for within-session predictions of each pain threshold measure was obtained by averaging diagonal values of the performance matrix in Fig. 6B, while performance for between-session predictions was obtained by averaging non-diagonal values of the matrix. We found that the performance of within-session predictions (heat pain on leg: r = 0.60; heat pain on arm: r = 0.56; pressure pain on leg: r = 0.57; pressure pain on thumbnail: r = 0.53) was slightly higher than between-session predictions (heat pain on leg: r = 0.56; heat pain on arm: r = 0.55; pressure pain on leg: r = 0.55; pressure pain on thumbnail: r = 0.50) for different pain threshold measures.

Fig. 6. Performance of predicting pain threshold within and between sessions.

A. Performance of within-session and between-session predictions. The performance for within-session predictions of each pain threshold measure was obtained by averaging diagonal values of the performance matrix, while performance for between-session predictions was obtained by averaging non-diagonal values of the performance matrix. B. The performance matrix shows the correlations between predicted and real values of pain threshold using data from different sessions as training and testing sets. R1–4: Resting state fMRI data 1–4; ns: no significant.

Fig. 6B shows the performance of within (diagonal) and between-session (non-diagonal) predictions in all pairs of training and testing sets, depicted by the performance matrix. Prediction models could reliably and significantly predict individual’s pain threshold scores across sessions in most pairs, especially for predicting heat pain threshold on the leg. The consistency was violated for the prediction of pressure pain threshold on the leg when using the data from the second session (R2: Resting 2) as the training set.

3.4. Effects of Head Motion

We did not observe any significant difference in head motion between the four sessions (F_3,69 = 0.94, p = 0.43), and individuals’ head motion was not correlated with their pain threshold in any session (p > 0.05 for all correlations).

Including the head motion as a feature in the prediction models could not increase the variation explained by the model without head motion, both for within-session predictions (p = 0.44, paired-sample t-test between 16 models, Fig. 6B diagonal values) and between-session predictions (p = 0.39, paired-sample t-test between 48 models, Fig. 6B non-diagonal values).

4. Discussion

In the present study, we found that resting-state functional brain connectivity is unique to each individual and encodes individual variability in pain thresholds. Test-retest analyses revealed that individual variability in connectivity as well as four pain threshold measures were both substantial across individuals and replicable within individuals. By using multivariate pattern analyses, we demonstrated that connectivity profiles can be used to predict individuals’ pain threshold scores at both the within-session level and between-session level. These results demonstrate the potential of using an fMRI brain connectome to build a ‘neural trait’ for characterizing an individual’s pain and pain-related behaviors, and this model may eventually be used to personalize clinical practices.

4.1. Reproducibility of Identifying Individuals Using the Brain Connectome

In a previous study, Finn et al. found that individuals can be identified by their connectivity matrices using data separated by a single day (Finn et al., 2015). They found that identification based on the whole-brain connectivity matrix was highly successful, and the combination of medial frontal and frontoparietal networks showed the highest inter-subject variance, resulting in the most discrimination between individuals. However, due to the neural plasticity of the brain, it remains unclear to what degree individual connectivity profiles are consistent across a longer period. Our study employed a longitudinal design over one month, and the cross-session identification rates were significantly higher than random. Those identification rates were comparable from data separated by one week, two weeks, and three weeks, indicating subjects’ stability in functional connectivity over the course of a month. Similar to the previous study, the medial frontal and frontoparietal networks showed the highest inter-subject variance and substantial within-subject stability. These two networks are comprised of higher-order association cortices. They have been suggested as playing a major role in large-scale coordination of brain activity (Cole et al., 2014; Power et al., 2011) and are also involved in cognitive control of the human brain (Dixon et al., 2018; Zanto and Gazzaley, 2013). With these findings (Finn et al., 2015), we uphold that the functional organization in the medial frontal and frontoparietal networks is highly variable during a task or at rest in individual subjects.

4.2. Association of Pain Threshold and Brain Connectome

Similar to the connectivity profiles, the behavioral results in the present study demonstrated that subjects had strong inter-subject variability in pain threshold while substantial within-subject stability across the four sessions, indicating personal pain trait. We assessed two pain modalities that are mediated by C-fibers (heat pain) and A-delta fibers (pressure pain) using quantitative sensory testing. We found a strong correlation in the two measures within each modality and a weaker correlation in the two measures between the modalities. Since they have different neural mechanisms, it is natural to suspect that individuals’ pain characteristics are different for heat pain and pressure pain, and such a difference may also extend to the patterns of neural trait for pain threshold.

Given the converging findings of inter-subject variability and within-subject stability of connectivity profiles and pain threshold, we explored whether individual differences in functional connectivity are relevant to individual differences in pain threshold. Cross validation within each session and external validation between sessions showed that connectivity profiles could be used to significantly predict subjects’ heat and pressure pain threshold scores. The medial frontal and frontoparietal networks also emerged as most predictive of pain threshold, which is consistent with previous studies showing that functional connectivity of the frontoparietal network could modulate cognitive aspects of pain (Kong et al., 2013; Lobanov et al., 2013). Disrupted functional connectivity in the medial frontal and frontoparietal networks have been linked to a variety of chronic pain conditions (Glass et al., 2011; Hemington et al., 2016; Kucyi and Davis, 2015; Napadow et al., 2010; Tu et al., 2019a), suggesting that the functional properties of these networks are important in modulating an individual’s pain and pain-related behaviors.

Although the medial frontal and frontoparietal networks were the most predictive, we observed different patterns of predictive edges for different pain modalities. For instance, we found that cerebellar functional connectivity played an important role in predicting heat pain threshold. Previous studies have suggested that the cerebellum may have a role in pain and nociceptive processing, and they showed evidence that nociceptive afferents project to the cerebellum (Borsook et al., 2008). In addition, the cerebellum has been associated with a number of different cortical areas involved in functional processes such as motor control, anticipation of pain, and negative emotions (Moulton et al., 2010). A large number of fMRI studies using heat pain stimuli show activation in the cerebellum (Apkarian et al., 2005; Borsook et al., 2008; Peyron et al., 2000), but only a few of these studies have found cerebellar activation using pressure pain (for a detailed review, see Moulton et al., 2010). In our results, connectivity from frontal, temporal, and parietal lobes made major contributions to the prediction of pressure pain. We speculate that the functional organization between these lobes predefines individuals’ cognitive and emotional brain states and therefore reflects aspects related to pain experience separated from sensory information.

A large number of neural markers in different aspects of pain have been identified, but the sensitivity of these markers may be compromised if inferences are drawn only at the group level due to inter-subject variability. For example, a pain prediction model trained on a group of training subjects may perform poorly on new subjects (Hu and Iannetti, 2016). Coghill et al. found that the anterior cingulate cortex, primary somatosensory cortex, and prefrontal cortex were the neural correlates of pain sensitivity after heat pain stimuli (Coghill et al., 2003). Another study from the same group showed that grey matter density in the posterior cingulate cortex, precuneus, intraparietal sulcus, and inferior parietal lobule were associated with pain sensitivity. In our case, we used different characteristics (pain threshold) and a longitudinal design to perform extensive test-retest analyses. Indeed, with the developments of the neuroimaging field, replicability is needed.

4.3. Assessing Pain Sensitivity in Patient Populations

Assessing pain sensitivity is important in clinical applications since different patients have different prognoses and may require different treatment. One proposed mechanism that may contribute to the development and maintenance of chronic pain is pain sensitization (Meints et al., 2019). Increased pain sensitivity in chronic pain patients as indicated by pain threshold increase compared to healthy subjects has been repeatedly found in studies (Edwards et al., 2011; Rebbeck et al., 2015; Zhang et al., 2015). In particular, patients with chronic opioid use had significantly elevated pain sensitivity than those without (Zhang et al., 2015). However, some populations may be associated with decreased pain sensitivity. For instance, clinical observations have shown that schizophrenia patients tend to report little or no pain under various painful conditions, and this pain insensitivity could lead to a higher risk of morbidity and mortality (Dworkin, 1994).

Though clinical and behavioral studies have shown altered pain sensitivity in patient populations, it remains unclear whether the observed difference occurs as a result of impaired subjective evaluation of pain (e.g., schizophrenia patients may have impaired communication) and whether the perception and experience of pain are altered as a result of abnormal brain functioning in cortical and subcortical regions involved in the transmission and processing of nociceptive information. Using the functional brain connectome and multivariate pattern analysis, our study holds the potential to uncover brain networks supporting pain hyper- or hyposensitivity and assess pain sensitivity when patients are unable to communicate pain effectively or when self-reports are otherwise suspect. However, our findings and the brain connectome-based model may be less accurate in patients and need to be further tested in a paradigm with patients.

4.4. Limitation and Conclusion

There are several limitations to the present study. First, the sample size is small. A recent study suggested that a cross-validation approach may lead to unstable and biased estimates when the sample size is small (Varoquaux et al., 2017). However, our within-subject design (each subject was evaluated four times) and extensive cross-session validation increased the power of the approach. Second, the problem of replicability also includes multisite validation. Further studies using data from an independent site is needed. Third, previous studies suggest that females may have altered pain sensitivity during menstrual cycles. Our present study did not record detailed information regarding menstrual phase. Future studies with large sample sizes may characterize the relationships between the identified brain connectome, menstrual cycle, and pain. Fourth, pain and neural activity change with age. Thus, it would be useful to study the stability or evolution of pain characteristics and neural markers across the years or even over lifespans.

In conclusion, our study for the first time applied a longitudinal design over one month with different pain measurements and found that strong individual differences in resting state functional connectivity are relevant to individual differences in pain sensitivity as measured by pain threshold. Examining how individuals’ networks are functionally organized in unique ways and relating this functional organization to pain sensitivity and other related behaviors are highly important for both healthy and patient populations and may facilitate the development of individualized medicine.