Expression of the neurologic pain signature during visceral and somatic stimulation (NPS generalizability)

We tested whether the NPS responded similarly to both aversive/painful visceral and somatic stimulation. To this end, we computed the expression of the NPS for contrasts of aversive stimulation versus nonpainful stimulation or aversive stimulation versus baseline (depending on the study; see “Methods” for details). NPS expression in the four visceral pain/discomfort studies is shown in Fig. 1a, with the highest response in the esophageal pain study (Study 5). NPS response was 0.043 ± 0.009 [t(14) = 4.66, P = 0.0004, d_a = 1.20 [0.423, 1.977], accuracy = 86.67%] for Study 1 (gastric pain); 0.045 ± 0.010 [t(14) = 4.69, P = 0.0003, d_a = 1.21 [0.432 1.988], accuracy = 93.33%] for Study 2 (rectal discomfort); 0.053 ± 0.007 [t(28) = 8.08, P < 0.0001, d_a = 1.50 [0.917, 2.083], accuracy = 96.6%] for Study 3 (rectal discomfort), and 0.122 ± 0.006 t(29) = 20.49, P < 0.0001, d_a = 3.74 [2.901, 4.579], accuracy = 100%] for Study 5 (esophageal pain). These effect sizes (d_a) are 1.5–5 times larger than typical “large effects”²⁷. The accuracy statistics reported above indicate that the NPS can accurately detect which of the two conditions is noxious for an individual participant in a forced-choice test.

Robust regression analysis demonstrated that the NPS response significantly predicted pain/discomfort intensity VAS ratings, even though the stimulus intensity was individually calibrated (r_weighted = 0.40, β_robust = 12.44 ± 04.45, P = 0.0064; Fig. 1a). Individual samples for visceral studies were correlated r = 0.38, r = 0.20, and r = 0.30 for gastric and the two rectal studies, respectively. We note that sample sizes varied from N = 15–29, and therefore individual-study correlations are likely to be highly unstable²⁸. A full treatment of individual differences in pain responses remains for future larger-sample studies.

The NPS response in the 3 somatic pain studies is shown in Fig. 1b. NPS response was 0.093 ± 0.011 [t(14) =  8.61, P < 0.0001, d_a = 2.22 [1.310, 3.130], accuracy = 100%] for Study 4 (vulvar pain); 0.132 ± 0.009 [t(27) = 15.22, P < 0.0001, d_a = 2.88 [2.132, 3.628], accuracy = 100%] for Study 6 (thermal pain) and 0.089 ± 0.007 [t(32) = 12.11, P < 0.0001, d_a = 2.11 [1.508, 2.712], accuracy = 100%] for Study 7 (thermal pain).

Overall, these results indicate that the NPS responded robustly to both somatic and visceral pain, and correlated with the subjective visceral pain experience. Comparing effect sizes, effects are larger in thermal, vulvar, and esophageal pain than rectal and gastric pain, suggesting that while the NPS generalizes, systems beyond the NPS may also be important for gastric and rectal pain.

Expression of other affective signatures during visceral and somatic stimulation (NPS sensitivity)

Because visceral pain is often thought to evoke stronger affect than somatic pain, we additionally tested whether brain responses to visceral and somatic stimulation were more similar to the NPS than other brain-based markers that track self-reported differences in social rejection²³, picture-induced negative emotion²², and vicarious pain²⁴.

As illustrated in Fig. 2a, overall studies together, data acquired during pain correlated significantly and positively with the NPS (average within-person Pearson spatial correlation r = 0.074 ± .0030 (SE), t(164) = 24.41, d_a = 1.92 [1.662, 2.177], P < 0.0001, q_FDR < 0.05). This association was significant for each individual study, across both somatic and visceral types (Fig. 2b). Turning to signatures for other affective processes in healthy individuals, data acquired during pain also correlated significantly and positively with the neural signatures for social rejection (r = 0.017 ± 0.0019, t(164) = 8.77, d_a = 0.697 [0.526, 0.866], P < 0.0001, q_FDR < 0.05) and negative emotion (PINES) (r = 0.006 ± 0.0019, t(164) = 3.45, d_a = 0.246 [0.091, 0.4004], P = 0.0007, q_FDR < 0.05), demonstrating some match to these other patterns, but with associations 4 and 12 times weaker, respectively, than associations with the NPS (for full details, see Supplementary Table S1). In addition, the positive associations with these patterns were not consistent across pain types; associations with the rejection pattern were driven by one rectal and the thermal studies, and associations with the PINES were driven by esophageal and vulvar studies (Fig. 2f). A significant negative correlation was found with the signature for vicarious pain (r = −0.012 ± 0.0021, t(164) = −5.78, d_a = −0.445 [−0.604, −0.284], P < 0.0001, q_FDR < 0.05), indicating lack of a positive match to this pattern. The correlation with the NPS was significantly stronger than the correlations with the three other neural signatures (ANOVA F(3,492) = 286.23, P < 0.0001; P < 0.0001 for all three pairwise comparisons versus NPS after Bonferroni correction for multiple testing).

a The similarity between brain responses pooled across pain types. Pearson correlations between pooled brain responses and neural signatures for somatic pain (NPS), vicarious pain (VPS), social rejection, and negative emotion (PINES). Lines show the mean association across participants, with shaded standard errors. The inner gray circle reflects a correlation of zero; points inside the circle are negative associations, and outside the circle are positive associations. Brain maps depict the 10,000 most positive and 10,000 most negative model weights on the cortical surface of the left hemisphere²⁹ and the basal ganglia³⁰. These weights are shown for display purposes only to indicate the brain regions that have the greatest influence in making predictions. Positive weights are shown in warm colors and negative weights in cool colors. *P < 0.05 (two-tailed, one-sample t test). b–e A visual depiction of the four neural signatures, with the same threshold as in (a). Warm colors indicate where increased activity leads to more positive outcomes and cool colors indicated where decreased activity leads to more positive outcomes. f Correlations between the average activity for each study. Scale and axes are identical to those in panel (a). *P < 0.05 (two-tailed, one-sample t test). Source data are provided as a Source Data file.

Expression of the neurologic pain signature during pain, negative emotion, and cognitive control (NPS specificity)

Because the studies used to examine commonalities between somatic and visceral pain in the analyses reported above differ in terms of scanning hardware, pulse sequences, participant demographics, and experimental designs, we conducted a separate validation to evaluate whether the NPS can reliably discriminate individual brain responses to painful stimuli from nonpainful control manipulations across a sample of independent studies (i.e., specificity of the NPS), with 6 studies manipulating pain (by thermal, mechanical, or rectal stimulation) and 12 control studies involving either manipulations of cognitive control or negative emotion (see ref. ³¹ for details). This analysis revealed that the NPS could effectively discriminate brain responses to different kinds of pain from conceptually related experimental manipulations (AUROC = 0.93, sensitivity = 73%, specificity = 92%, accuracy = 86 ± 2.1%, d_a = 2.13 [1.804, 2.449], Fig. 3).

Voxel-wise analysis of responses common to aversive somatic and visceral stimulation

To identify brain regions that exhibited similar changes in fMRI activation during somatic and visceral stimulation, we performed a voxel-based conjunction analysis of visceral and somatic stimulation modalities (each thresholded at q < 0.05 false discovery rate (FDR)-corrected³²). For each of the visceral (N = 89) and somatic subsets (N = 76), we controlled for the effect of the study using covariates for differences across studies (four visceral studies and three somatic studies, total N = 165, see “Methods” for details).

The conjunction analysis revealed that regions activated by somatic (Fig. 4a and Supplementary Table S2) and visceral stimulation (Fig. 4b and Supplementary Table S4) overlapped in a distributed set of brain regions (Fig. 4c and Supplementary Table S6), including midbrain, cerebellum, lentiform nucleus (putamen and pallidum), hypothalamus, thalamus (ventral lateral and ventral posterior lateral nucleus), parahippocampal gyrus/entorhinal cortex, insula (posterior, middle, and anterior parts), postcentral gyrus (including a medial cluster in primary somatosensory cortex (SI), and a ventral lateral cluster, including parietal/Rolandic operculum/secondary somatosensory cortex (SII)), adjacent inferior parietal lobule, superior temporal gyrus, and inferior frontal gyrus (ventrolateral prefrontal cortex (vlPFC)), lateral precentral gyrus, including primary motor cortex (MI) and premotor cortex, anterior and posterior midcingulate cortex (aMCC, pMCC) and adjacent medial frontal gyrus, and superior/middle frontal gyrus (dorsolateral prefrontal cortex (dlPFC)).

Parametric group-level maps of responses across participants and studies, controlling for interstudy differences. a The contrast of somatic stimulation versus baseline. Warm colors correspond to increased levels of activation and cool color deactivation (N = 76). b The contrast of visceral stimulation versus baseline (N = 89). c The conjunction of activation and deactivation maps for somatic and visceral stimulation (N = 165; the intersection of significant somatic and visceral results). d Contrasts of somatic (N = 76) and visceral pain (N = 89). In the top panel, effects are masked with the map of somatic pain > baseline for orange areas and visceral pain > baseline for yellow areas. Thus, the resulting maps show stimulus-related increases that are stronger for somatic (orange) or visceral (yellow). In these bottom panels, the difference map is masked with somatic pain < baseline for cyan areas and visceral pain < baseline for purple areas. Thus, the resulting maps show areas where stimulus-related deactivation occurs and is stronger in somatic pain (cyan), and deactivation occurs and is stronger in visceral pain (purple).

The conjunction analysis also revealed that regions deactivated by somatic (Fig. 4a and Supplementary Table S3) and visceral stimulation (Fig. 4b and Supplementary Table S5) overlapped in multiple brain regions (Fig. 4d and Supplementary Table S7), including thalamus (pulvinar), hippocampus, parahippocampal gyrus/perirhinal cortex, temporal pole, rostral middle/inferior temporal gyrus, occipital cortex and adjacent caudal middle and superior temporal gyrus, perigenual and subgenual anterior cingulate cortex (pACC, sACC) and adjacent medial, middle and superior frontal gyrus (vmPFC and dorsomedial prefrontal cortex (dmPFC)), lateral middle and superior frontal gyrus (dlPFC), PCC and adjacent precuneus, superior parietal lobule, left dorsal precentral gyrus (MI, premotor cortex), and left dorsal postcentral gyrus (SI).

Development and validation of a network-based classifier that differentiates visceral and somatic stimulation

To further test if representations of the different types of somatic and visceral aversive stimulations differ at a broader spatial scale, we next assessed whether brain responses to different types of visceral and somatic stimulation were differentially correlated with seven canonical resting-state cortical networks³³. This approach provides inferences about whether each of these networks is activated on average during each pain type. Detailed results of this analysis (based on point-biserial correlations) are provided as Supplementary Information (Supplementary Results, Supplementary Table S8). Broadly, they revealed some commonalities across all pain types, including activation of the “ventral attention” network and deactivation of the “default network” (Fig. 5a). (The network names are based on Yeo et al.³³, and by using them, we do not imply that their function is limited to or primarily related to the label). They also revealed differences: the “somatomotor” network was activated in some pain types (thermal, vulvar, and esophageal) but deactivated in others (rectal). The “frontoparietal” network was similarly activated by some (rectal, esophageal), but deactivated in others (thermal). Beyond these variations in which networks were activated and deactivated, there were also variations in the degree to which each network was activated or deactivated.

a A visual depiction of the resting-state networks based on data from 1000 subjects³³. b Point-biserial correlations between pooled brain responses (both somatic and visceral studies) and the seven resting-state networks. The inner bold line is the zero point, and values inside the inner circle reflect negative correlations. c Parameter estimates from a logistic regression model predicting whether brain responses were observed during visceral (rectal; N = 15) or somatic (cutaneous; N = 33) stimulation. Error bars reflect the standard error of the mean. *P < 0.05, two-tailed, one-sample t tests. d The Receiver Operating Characteristic (ROC) curve shows cross-validated performance on the training datasets (AU = area under the curve = 0.92). The solid line shows a Gaussian model fit. e Generalization tests show the probability of data from novel subjects as being classified as somatic or visceral. Scores for the training datasets were estimated using cross-validation. Test data scores apply the model prospectively, with no parameter adjustment. Each circle indicates the prediction for a single subject. Polar plots depict spatial similarity with resting-state networks as in panel (b). Each polar plot depicts the average similarity for each study: the rectal and thermal stimulation studies used for training the classifier, and the five hold-out studies used to test its generalizability. Source data are provided as a Source Data file.

The variations noted above could serve as the basis for a brain-based classifier for somatic versus visceral pain capable of making accurate predictions about new individual participants. Many pain types (e.g., esophageal, gastric, vulvar) may potentially include both somatic and visceral elements and cannot be defined as purely somatic or visceral a priori. Therefore, we first trained a logistic regression-based pattern classifier to discriminate between the clearest examples of somatic and visceral aversive stimulation available: cutaneous thermal stimulation (Study 7) and rectal distension with an inflatable balloon (Study 2). The data used for the classification were the spatial correlations of each individual participant’s stimulus-induced activity map with each of the seven resting-state networks. We tested classification accuracy on out-of-sample participants in Studies 2 and 7 using ten-fold cross-validation. Then, we tested the classifier’s performance on data for all pain types in independent studies, including thermal (Study 6) and rectal (Study 3) studies, and gastric (Study 1), vulvar (Study 4), and esophageal (Study 5) studies.

The classifier exhibited high levels of discriminability for thermal versus rectal when applied to out-of-sample participants in cross-validation (AUROC = 0.92, sensitivity = 94%, specificity = 80%, balanced accuracy = 87 ± 4.4%, d_a = 2.22 [1.46, 2.98]). As illustrated in Fig. 5b, c and Supplementary Table S8, somatic stimulation was indicated by a combination of (a) more positive correlations with the somatomotor network (logistic regression \((\hat{\upbeta}= 19.89 \pm 5.94, P=0.0008)\), driven by a positive correlation in thermal versus negative in rectal; (b) positive correlations with the dorsal attention network \((\hat{\upbeta}=14.21 \pm 5.72, P=0.013)\) driven by zero correlation in thermal versus a negative correlation in rectal; and (c) more positive correlations with the ventral attention network \((\hat{\upbeta}= 12.54 \pm 5.50, P=0.023)\), driven by a more strongly positive correlation in thermal. On the other hand, visceral stimulation was indicated by (a) more positive correlations with the frontoparietal network \((\hat{\upbeta}= −12.35 \pm 4.09, P = 0.0025)\), driven by a significant positive correlation in rectal versus a near-zero negative correlation in thermal; and (b) more positive correlations with the default network \((\hat{\upbeta}= −11.89 \pm 4.02, P = 0.0031)\), driven by a near-zero negative correlation in rectal versus a strong negative correlation in thermal. More positive correlations with the limbic network exhibited a nonsignificant trend towards predicting visceral stimulation (\(\hat{\upbeta}= −8.66, P = 0.063\), driven by a zero correlation in rectal versus a nonsignificant negative correlation in thermal).

Testing the pattern classification model on the remaining studies, which include gastric (Study 1), rectal (Study 3), vulvar (Study 4), esophageal (Study 5), and thermal (Study 6) stimulation, revealed a continuum of somatovisceral expression. The classifier demonstrated successful prospective generalization to new studies when tested on the rectal (Study 3) and thermal (Study 6) test sets (AUROC = 0.84, sensitivity = 82%, specificity = 69%, balanced accuracy = 76 ± 5.7%, d_a = 1.34 [0.765, 1.915]). The esophageal (P_somatic = 96.67 ± 3.28%, P < 0.0001) and vulvar (P_somatic = 86.67 ± 8.78%, P = 0.083) studies were largely classified as somatic (although the latter only exhibited a nonsignificant trend), whereas the gastric (P_somatic = 66.67 ± 12.17%, P = 0.97) study exhibited an intermediate classification rate (Fig. 5e).

Voxel-wise analysis of differential responses to aversive somatic and visceral stimulation

To identify brain regions that exhibited differential levels of fMRI activity for visceral and somatic stimulation, we conducted a standard voxel-wise general linear model (GLM) analysis comparing stimulation modalities while controlling for the effect of individual studies (four visceral studies and three somatic studies, total N = 165, see “Methods” for details).

Of the regions that showed significant activation (i.e., increased activity to somatic or visceral stimulation compared to baseline), regions demonstrating significantly stronger activation during somatic compared to visceral stimulation included midbrain, cerebellum, basal ganglia (caudate, lentiform nucleus, claustrum), thalamus (midline/medial dorsal/anterior nuclei), insula (posterior, middle, and anterior), inferior parietal lobule, postcentral gyrus/SI, parietal operculum/SII, temporoparietal junction, aMCC/pMCC, (pre)motor cortex, and middle frontal gyrus (dlPFC) (Supplementary Table S9 and Fig. 4a, b, e). Regions more strongly activated by visceral stimulation included thalamus (pulvinar), lentiform nucleus, amygdala, (caudal) parahippocampal gyrus, fusiform gyrus, occipital cortex (lingual gyrus, cuneus), precuneus/PCC, middle temporal gyrus, inferior and superior parietal lobule, and precentral/middle frontal gyrus (details in Supplementary Table S10 and Fig. 4a, b, e).

Of the regions that showed significant deactivation (i.e., decreased activity to somatic or visceral stimulation compared to baseline), regions demonstrating significantly stronger deactivation during somatic compared to visceral stimulation included rostral hippocampus/parahippocampal gyrus/entorhinal cortex, fusiform gyrus, temporal pole, middle and inferior temporal gyrus, occipital cortex (inferior & middle occipital gyrus, lingual gyrus, cuneus), pACC, sACC, vmPFC, orbitofrontal cortex (OFC) and dmPFC, PCC, precuneus, superior and inferior parietal lobule, vlPFC, and right dlPFC (Supplementary Table S11 and Fig. 4a, b, f). A more limited number of regions showed significantly stronger deactivation during visceral compared to somatic stimulation, including caudal hippocampus/parahippocampal gyrus and adjacent fusiform gyrus, caudate, right precentral/medial frontal gyrus, and right superior parietal lobule (Supplementary Table S12 and Fig. 4a, b, f).