Identifying brain systems in personalized and population-level predictive models of pain
Poster No:
2519
Submission Type:
Abstract Submission
Authors:
Youngeun Park1,2, Sungwoo Lee2,1, Dong Hee Lee2, Choong-Wan Woo1,2
Institutions:
1Sungkyunkwan University, Suwon, Korea, Republic of, 2Center for Neuroscience Imaging Research, Institute for Basic Science, Suwon, Korea, Republic of
First Author:
Youngeun Park
Sungkyunkwan University|Center for Neuroscience Imaging Research, Institute for Basic Science
Suwon, Korea, Republic of|Suwon, Korea, Republic of
Sungkyunkwan University|Center for Neuroscience Imaging Research, Institute for Basic Science
Suwon, Korea, Republic of|Suwon, Korea, Republic of
Co-Author(s):
Sungwoo Lee
Center for Neuroscience Imaging Research, Institute for Basic Science|Sungkyunkwan University
Suwon, Korea, Republic of|Suwon, Korea, Republic of
Center for Neuroscience Imaging Research, Institute for Basic Science|Sungkyunkwan University
Suwon, Korea, Republic of|Suwon, Korea, Republic of
Dong Hee Lee
Center for Neuroscience Imaging Research, Institute for Basic Science
Suwon, Korea, Republic of
Center for Neuroscience Imaging Research, Institute for Basic Science
Suwon, Korea, Republic of
Choong-Wan Woo
Sungkyunkwan University|Center for Neuroscience Imaging Research, Institute for Basic Science
Suwon, Korea, Republic of|Suwon, Korea, Republic of
Sungkyunkwan University|Center for Neuroscience Imaging Research, Institute for Basic Science
Suwon, Korea, Republic of|Suwon, Korea, Republic of
Introduction:
Pain is a complex and subjective experience that extends beyond a simple nociceptive process. Various psychological and social factors influence pain experiences and contribute to individual variability in brain representations of pain (Fillingim, 2017; Kohoutova et al., 2022). In attempts to develop clinically useful pain biomarkers, current group-level fMRI pain biomarkers have demonstrated common representations of neural responses to pain and shown abilities to predict pain across different (Wager et al., 2013; Woo et al., 2017). However, their clinical potential remains elusive. On the other hand, the precision neuroimaging approach enables to describe individual-specific brain organization through extensive sampling of single individuals (Gordon et al., 2017). Using this approach, personalized models may have the potential to obtain stable and reliable within-individual measures (Kraus et al., 2023). In this study, we developed a personalized and population-level pain predictive models, examined its capacity to predict pain within and across individuals, and identified similarities and differences in the important brain regions involved in pain prediction.
Methods:
A total of 125 healthy, right-handed participants participated in the study. One participant (author C.-W. W.) completed 33 sessions for a total of 66 hours of fMRI scanning for the individual dataset. The remaining 124 participants underwent a single session for the population dataset. During the fMRI scan, the participants experienced a series of thermal pain stimuli, ranging from 45 to 47.5℃ in 0.5℃ increments, and rated the intensity of pain induced by the stimulus across 96 trials per session.
In the development of pain predictive models, we randomly divided single-trial fMRI datasets for the individual and population data into training, validation, and test sets based on sessions. We trained each model to predict pain ratings using principal component regression (PCR) with leave-one-session-out cross-validation (fig. 1a) and tested the models in the independent test sets (fig. 1b).
To compare the weights of the two models, we measured searchlight-based pattern similarity, calculating the correlation coefficients between the predictive weights of the models within each searchlight. In the searchlight-based virtual lesion analysis, we excluded the predictive weights within searchlights, predicted pain ratings with the remaining predictive weights, and calculated the changes in the prediction performance.
In the development of pain predictive models, we randomly divided single-trial fMRI datasets for the individual and population data into training, validation, and test sets based on sessions. We trained each model to predict pain ratings using principal component regression (PCR) with leave-one-session-out cross-validation (fig. 1a) and tested the models in the independent test sets (fig. 1b).
To compare the weights of the two models, we measured searchlight-based pattern similarity, calculating the correlation coefficients between the predictive weights of the models within each searchlight. In the searchlight-based virtual lesion analysis, we excluded the predictive weights within searchlights, predicted pain ratings with the remaining predictive weights, and calculated the changes in the prediction performance.
Results:
The personalized model exhibited the best predictive performance when predicting pain rating within the individual (mean r = 0.774) compared to the population model testing on the population data (mean r = 0.672) (Fig. 1b left). The two-sample t-test result showed significant differences between the prediction performances (Fig. 1b right). However, both models also could predict pain in the other dataset, indicating their utility as pain-predictive models. The weight patterns of the two models were highly similar in the regions including insula and supplementary motor cortex, while the regions including lateral thalamus and vmPFC showed low similarities (Fig. 1c & Fig. 1d). In the virtual lesion analysis, broader brain regions were important in predicting pain within the individual using the personalized model (one→one) compared to the population model (pop→one). The insula, sensorimotor areas, and high-order regions (dmPFC and dlPFC) were uniquely important in one→one (Fig. 1f).
Conclusions:
Overall, we showed that the personalized pain predictive model outperformed the group-level model for individual-specific pain prediction and exhibited importance in the high-order areas and the insula, which are responsible for pain processing and modulation, particularly cognitive and emotional aspects of pain experiences. These findings highlight the importance of developing personalized neuroimaging-based pain biomarkers, supporting personalized pain assessment and treatment.
Modeling and Analysis Methods:
Classification and Predictive Modeling 2
Novel Imaging Acquisition Methods:
BOLD fMRI
Perception, Attention and Motor Behavior:
Perception: Pain and Visceral 1
Keywords:
FUNCTIONAL MRI
Modeling
Pain
1|2Indicates the priority used for review
Provide references using author date format
Kohoutová, L., Atlas, L. Y., Büchel, C., Buhle, J. T., Geuter, S., Jepma, M., ... & Woo, C. W. (2022). Individual variability in brain representations of pain. Nature neuroscience, 25(6), 749-759.
Wager, T. D., Atlas, L. Y., Lindquist, M. A., Roy, M., Woo, C. W., & Kross, E. (2013). An fMRI-based neurologic signature of physical pain. New England Journal of Medicine, 368(15), 1388-1397.
Woo, C. W., Schmidt, L., Krishnan, A., Jepma, M., Roy, M., Lindquist, M. A., ... & Wager, T. D. (2017). Quantifying cerebral contributions to pain beyond nociception. Nature communications, 8(1), 14211.
Gordon, E. M., Laumann, T. O., Gilmore, A. W., Newbold, D. J., Greene, D. J., Berg, J. J., ... & Dosenbach, N. U. (2017). Precision functional mapping of individual human brains. Neuron, 95(4), 791-807.
Kraus, B., Zinbarg, R., Braga, R. M., Nusslock, R., Mittal, V. A., & Gratton, C. (2023). Insights from Personalized Models of Brain and Behavior for Identifying Biomarkers in Psychiatry. Neuroscience & Biobehavioral Reviews, 105259.