A multistudy analysis reveals that evoked pain intensity representation is distributed across brain systems

Figures

• 

  Fig 1. Study design. (A) Each experimental design involved a succession of thermal boxcar stimuli delivered at variable but finite intervals over a succession of fMRI sessions. A psychological manipulation designed to affect pain intensity preceded the stimulus in many designs, for example, a pain-predictive cue designed to heighten or reduce pain experience. Thermal stimuli varied in intensity and were mostly delivered to the left ventral forearm as illustrated. Cues varied across studies and were uncorrelated with stimulus intensity. After each stimulus, participants rated their pain intensity using a visual scale. For details on each design, refer to Table 2. (B) Pain ratings reflect a variety of influences on pain experience. Wedges show the proportion of variance explained by each factor for each study individually and across all studies (excluding validation studies, balancing participants across studies). Since fMRI data were quartiled for computational tractability, and z-scored for between study normalization, these data were also quartiled and (in the aggregated data) z-scored prior to regression analyses. The sum of wedges equals model variance explained. Residual model error…

• 

  Fig 2. We tested 6 different hypotheses regarding the scope of pain representation by inspecting the decoding performance of models trained on each of 6 corresponding sets of brain areas. Three modular sets of areas (top row) tested predictive models based on single regions and networks. A single most predictive “module”—region, fRSN, or cRSN—was chosen, and voxel weights within the module were optimized to predict pain. The modules are shown in different colors. Three additional nonmodular sets of areas (bottom row) tested if the BOLD signal across sets of distributed and functionally heterogeneous areas best represented pain. All performance metrics were evaluated using nested CV to ensure independent testing, training, and (where applicable) module selection. Underlying atlas and mask data: https://github.com/canlab/petre_scope_of_pain_representation/tree/main/figure2. BOLD, Blood Oxygen Level Dependent; cRSN, coarse resting-state network; CV, cross-validation; fRSN, fine resting-state network.

• 

  Fig 3. Multiple brain areas produced models with significant pain decoding capabilities, with prediction accuracy varying considerably across participants (error bars show SEM; n = 112). Predictions from 51 elementary regions, 7 fine-scale (fRSN) and 4 coarse-scale (cRSN) resting-state networks correlated significantly with pain reports (shown; ɑ = 0.05, twice repeated 5-fold cross-validated predictions modeled with random participant and random study effects, Šidák correction for multiple comparisons, effective p < 10e-4). This means the “best” region, cRSN or fRSN, is highly dependent on the participant sample examined (and will differ across folds in a CV scheme). Regions are color-coded by large-scale structures (i.e., subdivisions of Ctx, Cblm, Bstem, or Thal). cRSNs and fRSNs are colored by individual network parcels. Underlying data: https://github.com/canlab/petre_scope_of_pain_representation/tree/main/figure3. Cblm, cerebellum; cRSN, coarse resting-state network; Ctx, cortex; CV, cross-validation; fRSN, fine resting-state network; Thal, thalamus.

• 

  Fig 4. Multivariate brain models of pain response can be interpreted as nested models that draw on increasing amounts of unique information. Which module—single-region, fine-scale (fRSN) and coarse-scale (cRSN) resting-state network—was selected by the algorithm varied across folds. Bar graphs illustrate frequency of area selection across 10 folds (2 × 5 repeated CV). Variability across folds is consistent with the large number of highly predictive regions and network parcels and the fact that predictive signal is distributed across them. Bootstrap-corrected maps were computed for all regions to characterize pattern stability. For modular spaces, maps from different CV folds were conjoined after bootstrap correction (i.e., the union of maps across folds), and we calculated pairwise correlations among weights from different models. Model weights are shown on representative brain slices, with yellow/orange indicating positive weights (pro-pain) and blue negative weights (antipain), thresholded with 5,008 bootstrap samples, voxel-wise p < 0.05, uncorrected. While the relatively small number of significant weights suggests low model stability, the pattern spatial similarity matrix…

• 

  Fig 5. Multisystem and full brain representations of pain performed better than modular representations of pain in a 2 × 5 repeated CV test of model performance. We Fisher z-transformed within-participant correlation coefficient (predicted versus observed) to quantify performance. (A) All obtained models showed significant predictive performance (p < 0.01, random participant and study intercept, random study effect), and the specific models obtained here to predict pain using distributed sets of brain areas outperformed the models trained on individual modules (planned comparison, mean(PainPathway, Neurosynth, FullBrain) − mean(regions, fRSN, cRSN) = 0.272(0.07), F1, 6.9 = 15.1, p = 0.0062). We confirmed the nullity of remaining planned contrasts (BF > 10 favoring the null), except for cRSN > fRSN about which we could reach no conclusion (BF = 5.3 in favor of cRSN = fRSN). Post hoc comparison showed a full brain model outperformed the best single region (FullBrain − Regions = 0.34(0.10), F1,7.9 = 12.39, p = 0.008). Models were trained after pooling data across 7 datasets, yielding 90 participants per training fold. (B) We retrained cross-validated models separately for each study to…

• 

  Fig 6. Full brain models captured brain correlates of both sensory (red, orange) and psychological (blue) influences on pain reports. Five experimental factors showed significant effects on pain reports: temperature, habituation, perceived control (control), expectation, and social cues (multivariate regression, fixed participant effects, p < 0.05). However, the full brain CV model predictions also significantly correlated with pain outcome (scatter plot), and temperature, habituation, and expectation, in turn, significantly correlated with these full brain model predictions (α, standardized multivariate regression coefficients; Model a pie chart show partial-r2, color-coded by factor). Finally, the full brain model continued to significantly predict pain outcome (β) even while controlling for experimental factors (τ). This suggested that temperature, habituation, and expectation effects were captured by the full brain model. Multivariate mediation analysis supported this showing both significant indirect (mediated, α*β), and direct (τ) effects on pain report, with additional significant direct effects from perceived control and social cues. Together, these factors and the full brain…

• 

  Fig 7. Distributed representations outperformed modular representations in validation datasets, consistent with estimated generalization performances. (A) All models fit to the 7 study dataset significantly predicted pain in novel data, except for 5 models in 5/28 cases (p < 0.05, left, only nonsignificant results labeled, Holm–Sidak corrected). Performance in validation data continued to be significantly better for multisystem and full models relative to modular models in validation studies 1 and 2 (mixed effects model, random participant intercept, F1,160 = 12.6, p < 10e-3; F1,220 = 11.8, p < 10e-3; resp.) as well as overall (mixed effects model, random participant intercept, and fixed study effects, F1,1005 = 15.6, p < 1e-4); however, the full brain model did not show better performance than the best elementary region model specifically (region = full brain, BF = 2.2e5; right). (B) Nevertheless, retraining all 6 families of algorithms in novel data reproduced both findings: full brain and multisystem representations predicted more pain than regional and network representations and the full brain model specifically out performed any single best region. In a 2 × 5 repeated CV scheme,…

• 

  Fig 8. Distributed models were more pain-specific than local models. In a set of 18 studies that estimated participant mean responses to 1 of 6 multimodal pain tasks or 1 of 12 nonpain tasks, models significantly discriminated pain from nonpain tasks on average (p < 10e-6), even though nonpain tasks all shared aversive characteristics or task demands with pain. Planned contrasts showed this discrimination was statistically significant for each of the 6 models tested. Notably, all models were exclusively trained on painful stimuli, so the ability to discriminate was an emergent phenomenon. Most importantly, discrimination of pain from nonpain was greater for distributed than local models (p = 2.2e-5), indicating that the performance advantage of distributed models was due to improved pain-specific representations. Bars are grouped into pain tasks (blue), cognitive demand control tasks (red), and nonpainful aversive tasks (yellow). Planned contrasts: black brackets, p < 0.003 (Sidak threshold for 17 comparisons), gray/black brackets p < 0.05. Post hoc t test: *p < 0.003. Dashed gray line: optimal cut point for discrimination (balancing sensitivity and specificity). Legend indicates…