Geometrical Features of Neural Trajectory as Computational Motifs for Pain Cue-Stimulus Integration
Poster No:
2493
Submission Type:
Abstract Submission
Authors:
Jungwoo Kim1,2,3, Suhwan Gim1,2, Seng Bum Yoo1,2, Choong-Wan Woo1,2,3
Institutions:
1Center for Neuroscience Imaging Research, Institute for Basic Science, South Korea, 2Department of Biomedical Engineering, Sungkyunkwan University, South Korea, 3Department of Intelligent Precision Healthcare Convergence, Sungkyunkwan University, South Korea
First Author:
Jungwoo Kim
Center for Neuroscience Imaging Research, Institute for Basic Science|Department of Biomedical Engineering, Sungkyunkwan University|Department of Intelligent Precision Healthcare Convergence, Sungkyunkwan University
South Korea|South Korea|South Korea
Center for Neuroscience Imaging Research, Institute for Basic Science|Department of Biomedical Engineering, Sungkyunkwan University|Department of Intelligent Precision Healthcare Convergence, Sungkyunkwan University
South Korea|South Korea|South Korea
Co-Author(s):
Suhwan Gim
Center for Neuroscience Imaging Research, Institute for Basic Science|Department of Biomedical Engineering, Sungkyunkwan University
South Korea|South Korea
Center for Neuroscience Imaging Research, Institute for Basic Science|Department of Biomedical Engineering, Sungkyunkwan University
South Korea|South Korea
Seng Bum Yoo
Center for Neuroscience Imaging Research, Institute for Basic Science|Department of Biomedical Engineering, Sungkyunkwan University
South Korea|South Korea
Center for Neuroscience Imaging Research, Institute for Basic Science|Department of Biomedical Engineering, Sungkyunkwan University
South Korea|South Korea
Choong-Wan Woo
Center for Neuroscience Imaging Research, Institute for Basic Science|Department of Biomedical Engineering, Sungkyunkwan University|Department of Intelligent Precision Healthcare Convergence, Sungkyunkwan University
South Korea|South Korea|South Korea
Center for Neuroscience Imaging Research, Institute for Basic Science|Department of Biomedical Engineering, Sungkyunkwan University|Department of Intelligent Precision Healthcare Convergence, Sungkyunkwan University
South Korea|South Korea|South Korea
Introduction:
The brain integrates information from pain-predictive cues and noxious inputs to construct pain (Wiech, K, 2016). While previous studies have identified the neural correlates of individual pain components, the computational mechanisms of their integration remain elusive. Here, we used human fMRI to investigate neural trajectories within subspaces that preserve cue and stimulus information in network regions. We utilized the geometrical features of these trajectories to elucidate how the brain preserves and integrates information about pain components. We also explored the integration mechanism in large-scale networks aligned at the cortical hierarchy (Margulies et al, 2016), investigating whether progressive integration occurs from unimodal networks to transmodal networks. Our results show that while all networks preserve information about individual pain components in their respective subspaces, only the transmodal networks integrate it through the linear summation of information from subspaces.
Methods:
In the fMRI scanner, we presented 56 participants with a 2-second visual pain-predictive cue (high, low, and no pain cue), followed by a heat stimulation at one of five intensity levels for 12.5 seconds (Fig. 1a). Participants then rated their pain experience.
Using temporal fMRI data during participants' pain experiences, we applied targeted dimensionality reduction (Mante et al, 2013) to separate the subspace accounting for cue and stimulus information in each network region (Fig. 1f). We then projected the data onto each subspace and measured the distance between the neural trajectories of pain conditions. We used it to measure how well the subspace encodes the corresponding information (Fig. 1d, g) through a generalized linear model fit. We also calculated cue and stimulus effects by subtracting the encoding performances of the null subspace from the encoding performances of the subspace of interest to assess their significance. (Fig. 1d).
We then linearly summed the distance information from each subspace and compared it with actual pain ratings (Fig. 1e) to investigate whether the network integrates the two types of information. Given the different scales for reconstructed and actual pain reports, we normalized the values after subtracting the middle condition value (i.e., the condition of no cue with stimulus intensity level 3) and dividing them by their variance. We measured the reconstruction fit by calculating the absolute difference between the two.
Using temporal fMRI data during participants' pain experiences, we applied targeted dimensionality reduction (Mante et al, 2013) to separate the subspace accounting for cue and stimulus information in each network region (Fig. 1f). We then projected the data onto each subspace and measured the distance between the neural trajectories of pain conditions. We used it to measure how well the subspace encodes the corresponding information (Fig. 1d, g) through a generalized linear model fit. We also calculated cue and stimulus effects by subtracting the encoding performances of the null subspace from the encoding performances of the subspace of interest to assess their significance. (Fig. 1d).
We then linearly summed the distance information from each subspace and compared it with actual pain ratings (Fig. 1e) to investigate whether the network integrates the two types of information. Given the different scales for reconstructed and actual pain reports, we normalized the values after subtracting the middle condition value (i.e., the condition of no cue with stimulus intensity level 3) and dividing them by their variance. We measured the reconstruction fit by calculating the absolute difference between the two.
Results:
Among the possible behavioral results (Fig. 1b), our behavioral results support the case of cue-stimulus integration, as evidenced by the significant effects of both the cue and stimulus intensity on the pain ratings (Fig. 1c).
For the encoding performances of the subspaces, all seven large-scale networks exhibited significant cue and stimulus effects (Fig. 2a), implying that all network regions preserve both types of information in their respective subspaces. Example low-dimensional trajectories of the visual network and the limbic network are shown in Figs. 2b and c, respectively, from which the encoding performances were calculated.
The linear summation of the distance measures from each subspace (Fig. 2d) indicated that the transmodal networks (the frontoparietal, limbic, and default networks) showed significant reconstruction fit (Figs. 2g-j), while the unimodal networks (the visual and somatomotor networks) did not (Figs. 2e-f, j), implying that only the transmodal networks integrate the information.
For the encoding performances of the subspaces, all seven large-scale networks exhibited significant cue and stimulus effects (Fig. 2a), implying that all network regions preserve both types of information in their respective subspaces. Example low-dimensional trajectories of the visual network and the limbic network are shown in Figs. 2b and c, respectively, from which the encoding performances were calculated.
The linear summation of the distance measures from each subspace (Fig. 2d) indicated that the transmodal networks (the frontoparietal, limbic, and default networks) showed significant reconstruction fit (Figs. 2g-j), while the unimodal networks (the visual and somatomotor networks) did not (Figs. 2e-f, j), implying that only the transmodal networks integrate the information.
Conclusions:
We have shown that all networks preserve information about pain expectations and noxious inputs. However, integration occurs only within the transmodal networks, through the linear summation of the information from each subspace. We demonstrated how distances between neural trajectories in networks' subspaces encode information about different pain components and could serve as a computational motif for eliciting actual pain behavior.
Modeling and Analysis Methods:
Activation (eg. BOLD task-fMRI)
Multivariate Approaches
Perception, Attention and Motor Behavior:
Perception: Multisensory and Crossmodal 1
Perception: Pain and Visceral 2
Keywords:
Computational Neuroscience
Multivariate
Pain
1|2Indicates the priority used for review
Provide references using author date format
Margulies et al., (2016), 'Situating the default-mode network along a principal gradient of macroscale cortical organization', Proceedings of the National Academy of Sciences, 113(44), 12574-12579.
Mante et al., (2013). 'Context-dependent computation by recurrent dynamics in prefrontal cortex', Nature, 503(7474),