Geometrical Features of Neural Trajectory as a Computational Motif for the Cue-Stimulus Integration of Pain

Pain is constructed in the brain by integrating multimodal information, ranging from noxious inputs to expectations. Previous studies have identified brain systems that mediate the effects of different components of pain, but the computational mechanisms of their integration have been understudied. Here we examined whether the neural trajectories of different component subspaces would reveal the computational principles of information integration. To this end, we conducted a cue-induced pain expectation experiment with functional Magnetic Resonance Imaging in humans (N = 56) to examine the computational mechanisms of the integration of cue and stimulus information in the brain. Applying a targeted linear dimensionality reduction method on the voxel-level dynamics in large-scale functional brain networks, we derived subspaces of cue and stimulus and examined the decoding performance for cue and stimulus information in each subspace and in each large-scale network. If the decoding performances of cue and stimulus information were high in both subspaces, we considered that the level of integration was high, and if the decoding performance was high only in one subspace, we considered the level of integration was low. We found that the unimodal brain networks, such as visual and somatomotor networks, showed modality-specific decoding performances (i.e., the visual network for the cue information and the somatomotor network for the stimulus information), suggesting a low level of integration. However, the limbic network (which consists of transmodal brain regions) showed high decoding performances for both cue and stimulus information, suggesting a high level of information integration. We also found that the dynamics from each subspace from the limbic network reconstructed pain ratings most accurately. Overall, this study suggests a possible computational mechanism for the integration of cue and stimulus information in the brain and provides a general framework for studying multimodal integration from the dynamic systems perspective.