Decoding and prediction of individual neural representations with a scalable tensor-based principle
Poster No:
1510
Submission Type:
Abstract Submission
Authors:
Shitong Xiang1, Rongquan Zhai1, Gunter Schumann1, JianFeng Feng1, Tianye Jia1
Institutions:
1Institute of Science and Technology for Brain-Inspired Intelligence, Fudan University, Shanghai, China
First Author:
Shitong Xiang
Institute of Science and Technology for Brain-Inspired Intelligence, Fudan University
Shanghai, China
Institute of Science and Technology for Brain-Inspired Intelligence, Fudan University
Shanghai, China
Co-Author(s):
Rongquan Zhai
Institute of Science and Technology for Brain-Inspired Intelligence, Fudan University
Shanghai, China
Institute of Science and Technology for Brain-Inspired Intelligence, Fudan University
Shanghai, China
Gunter Schumann
Institute of Science and Technology for Brain-Inspired Intelligence, Fudan University
Shanghai, China
Institute of Science and Technology for Brain-Inspired Intelligence, Fudan University
Shanghai, China
JianFeng Feng
Institute of Science and Technology for Brain-Inspired Intelligence, Fudan University
Shanghai, China
Institute of Science and Technology for Brain-Inspired Intelligence, Fudan University
Shanghai, China
Tianye Jia
Institute of Science and Technology for Brain-Inspired Intelligence, Fudan University
Shanghai, China
Institute of Science and Technology for Brain-Inspired Intelligence, Fudan University
Shanghai, China
Introduction:
Decoding neural representations of cognitive processes from complex fMRI signals remains a fundamental challenge. While various decomposition approaches exist (Kober et al., 2008; Xiang et al., 2023), they often rely heavily on experimental design, limiting analyses to a few specific contrasts and lacking multidimensional characterization (Smith & Nichols, 2018). Recent advancements in tensor decomposition of BOLD signals have incorporated spatiotemporal characterization (Li et al., 2021), but they still struggle with individualized features and aligning different time points. On the other hand, longitudinal experimental designs have been re-emphasized (Kang et al., 2024). However, existing tensor-based methods face difficulties when task conditions vary across time and individuals. Here, we introduce a novel tensor-decomposition-based analytical framework to decode emotional face processing at the task-condition levels.
Methods:
Total 618 participants from IMAGEN project, who completed the emotional face task at ages 14, 19, and 23, were included following quality control. After preprocessing (by fMRIprep) or first-level analysis (by SPM), the data were organized into a 3D matrix (participants × voxel number × time-series or task conditions). Using tensor Canonical Polyadic (CP) decomposition, we extracted three vectors representing individual process-specific scores, neural representations, and temporal or conditional loadings across the three dimensions (Fig. 1a).
Results:
At the time-series level, among the top ten components from the tensor decomposition, only the third component demonstrated cross-time stability in both the individual and spatial dimensions. Its neural representation exhibited the highest similarity to traditional contrast-based activation maps, hence indicating a meaningful tensor decomposition (Fig. 1b). At the task-condition level, only the first component showed cross-time stability. Furthermore, its neural representation closely resembled the third component identified in the time-series decomposition (Fig. 1c), supporting the feasibility of task-condition-based tensor decomposition. Interestingly, this neural representation could capture a comprehensive neural representation of emotional processing, given high similarity with all traditional activation maps (Fig. 1d). The individual-specific scores derived from this component also showed stable associations with depressive symptoms (Fig. 1e).
Notably, the happy face stimuli were introduced only during the later follow-up sessions. We extended our analytical principle to longitudinal data to impute the missing neural response at baseline. Specifically, we extracted the common parts across the time points and reorganized them into a 4D matrix, which was then subjected to CP decomposition. The meaningful components derived from this decomposition were subsequently incorporated into tensor regression, enriched with additional task conditions, to estimate the condition-specific loadings. Finally, the missing neural representations were predicted by computing the cross product between the shared components and the condition-specific weights obtained from the tensor regression (Fig. 2a). In general, the predicted neural responses at baseline showed strong individual specificity (Fig. 2b). At the group level, activation maps produced with contrast-based analysis on the predicted neural responses showed high spatial similarity to those derived from the actual follow-up data (Fig, 2c). More importantly, predictions incorporating prior information achieved higher accuracy, highlighting the scalability and robustness of our approach (Fig. 2d).
Notably, the happy face stimuli were introduced only during the later follow-up sessions. We extended our analytical principle to longitudinal data to impute the missing neural response at baseline. Specifically, we extracted the common parts across the time points and reorganized them into a 4D matrix, which was then subjected to CP decomposition. The meaningful components derived from this decomposition were subsequently incorporated into tensor regression, enriched with additional task conditions, to estimate the condition-specific loadings. Finally, the missing neural representations were predicted by computing the cross product between the shared components and the condition-specific weights obtained from the tensor regression (Fig. 2a). In general, the predicted neural responses at baseline showed strong individual specificity (Fig. 2b). At the group level, activation maps produced with contrast-based analysis on the predicted neural responses showed high spatial similarity to those derived from the actual follow-up data (Fig, 2c). More importantly, predictions incorporating prior information achieved higher accuracy, highlighting the scalability and robustness of our approach (Fig. 2d).
Conclusions:
Our novel tensor decomposition-based analytical framework, especially the imputations for the missing neural responses could advance the modelling of cognitive processes and facilitate brain-wide association studies. We anticipate our novel analytical principle to be applied generally in decoding latent neurobehavioral processes.
Emotion, Motivation and Social Neuroscience:
Emotional Perception 2
Modeling and Analysis Methods:
Methods Development 1
Keywords:
Cognition
Design and Analysis
Emotions
FUNCTIONAL MRI
Other - Tensor decomposition
1|2Indicates the priority used for review
Abstract Information
By submitting your proposal, you grant permission for the Organization for Human Brain Mapping (OHBM) to distribute your work in any format, including video, audio print and electronic text through OHBM OnDemand, social media channels, the OHBM website, or other electronic publications and media.
The Open Science Special Interest Group (OSSIG) is introducing a reproducibility challenge for OHBM 2025. This new initiative aims to enhance the reproducibility of scientific results and foster collaborations between labs. Teams will consist of a “source” party and a “reproducing” party, and will be evaluated on the success of their replication, the openness of the source work, and additional deliverables. Click here for more information. Propose your OHBM abstract(s) as source work for future OHBM meetings by selecting one of the following options:
Please indicate below if your study was a "resting state" or "task-activation” study.
Healthy subjects only or patients (note that patient studies may also involve healthy subjects):
Was this research conducted in the United States?
Were any human subjects research approved by the relevant Institutional Review Board or ethics panel? NOTE: Any human subjects studies without IRB approval will be automatically rejected.
Were any animal research approved by the relevant IACUC or other animal research panel? NOTE: Any animal studies without IACUC approval will be automatically rejected.
Please indicate which methods were used in your research:
For human MRI, what field strength scanner do you use?
Which processing packages did you use for your study?
Provide references using APA citation style.
[2] Xiang, S. , Jia, T. , Xie, C. , Zhu, Z. , Cheng, W. , & Schumann, G. , et al. (2023). Fractionation of neural reward processing into independent components by novel decoding principle. NeuroImage, 284, 120463.
[3] Smith, S. M. , & Nichols, T. E. . (2018). Statistical challenges in "big data" human neuroimaging. Neuron, 97(2), 263.
[4] Li, J., Wisnowski, J. L., Joshi, A. A., & Leahy, R. M. (2021). Robust brain network identification from multi-subject asynchronous fMRI data. NeuroImage, 227, 117615.
[5] Kang, K., Seidlitz, J., Bethlehem, R.A.I. et al. Study design features increase replicability in brain-wide association studies. Nature. https://doi.org/10.1038/s41586-024-08260-9