Mapping the Brain's Core Network for High-Order Cognition via Machine Learning

1715 

Abstract Submission 

guowei wu^1,2, Xiuyi Wang^1,2, Yi Du^1,2,3

¹CAS Key Laboratory of Behavioral Science, Institute of Psychology, Chinese Academy of Sciences, Beijing, China, ²Department of Psychology, University of Chinese Academy of Sciences, Beijing, China, ³Chinese Institute for Brain Research, Beijing, China

Guowei Wu  
CAS Key Laboratory of Behavioral Science, Institute of Psychology, Chinese Academy of Sciences|Department of Psychology, University of Chinese Academy of Sciences
Beijing, China|Beijing, China

Xiuyi Wang  
CAS Key Laboratory of Behavioral Science, Institute of Psychology, Chinese Academy of Sciences|Department of Psychology, University of Chinese Academy of Sciences
Beijing, China|Beijing, China

Yi Du  
CAS Key Laboratory of Behavioral Science, Institute of Psychology, Chinese Academy of Sciences|Department of Psychology, University of Chinese Academy of Sciences|Chinese Institute for Brain Research
Beijing, China|Beijing, China|Beijing, China

Understanding the role of functional connectivity (FC) between brain regions in facilitating complex cognitive functions is a fundamental question in the field of neuroscience. Researchers have shown that machine learning models can predict cognitive abilities by utilizing FC with moderate accuracy (Kong et al., 2021a, 2023b; Tian & Zalesky, 2021). However, these models have primarily focused on specific cognitive behaviors, failing to comprehensively reveal how FC networks support the complexity of human cognition. Although evidence suggests that FC networks exhibit higher accuracy in predicting cognitive ontology than single-task performance (Dubois et al., 2018), it remains uncertain whether specific subsets of FC, crucial for predicting cognitive ontology, can effectively represent various high-order cognitive functions. Furthermore, understanding the relationship between the FC connectome, structural connectivity and gene expression patterns is crucial for unraveling the mechanisms underlying cognitive processing.

To investigate the FC underpins of cognitive functions, we developed a machine learning framework using HCP Young Adult (HCP-YA) and Lifespan Development datasets. We derived a cognitive ontology factor from HCP-YA data and identified the key FC edges via ridge regression. Our proposed 'Cognitive Ontology Prediction Model' (COMP), comprising the most influential 10% of edges, was compared against two models - a full FC model (F-F) and a ridge regression-based connectome predictive model (rCPM) with the same number of edges as COMP. We evaluated the predictive accuracy of these models across 25 behavioral tasks and calculated the 'cognitive loading' through Spearman correlations between ontology score and behavioral scores. To examine whether the COPM model more precisely predict behaviors closely related to the cognitive ontology score compared to the F-F model and rCPM, we computed the Spearman correlation between cognitive loading and the prediction accuracy for the 25 behavioral metrics for each model. We then validated the COPM model on the HCP-D dataset to assess its predictive accuracy for high-order cognitive functions across different dataset. Furthermore, we examine the association between the FC edges' coefficients and biological markers like FC strength variability and gene co-expression. The robustness of our approach was confirmed by replicating the analyses using different brain atlas parcellations and edge selection criteria.

Using a second-order CFA model on HCP-YA dataset tasks, we derived a cognitive ontology factor emcompassing diverse cognitive functions, revealing a robust hierarchical structure (CFI = 0.95, RMSEA = 0.07). We then assessed the prediction accuracy of FC edges associated with this factor using ridge regression on the Schaefer 400 atlas-based FC matrix, achieving a median accuracy of 0.39. Furthermore, we correlated the cognitive loading of 25 behavioral tasks with prediction accuracy in the COPM, F-F, and rCPM models. Notably, COPM exhibited a superior correlation (r = 0.64, p < 0.001), significantly outperforming F-F and rCPM models. The validation of the COPM's predictive performance using the HCP-D dataset further substantiated its efficacy, particularly for high-order cognitive tasks. Additionally, we found a significant positive correlation between FC strength variability, white matter connectivity, gene co-expression, and absolute beta weights from the cognitive ontology prediction model. These edges predominantly resided within association networks like the Default Mode and Frontal-Parietal Networks. Our replication study using the Glasser 360 atlas and various edge selection thresholds reaffirmed the robustness of our findings.

Our investigation delved into the FC networks underpinning cognitive ontology and their broader implications for high-order cognitive functions. Brain's cognitive task reliance is influenced by structural connectivity, and genetics.

Higher Cognitive Functions Other ²

fMRI Connectivity and Network Modeling ¹

Methods Development

Cognition

Computational Neuroscience

FUNCTIONAL MRI

Machine Learning