Can function be predicted from structure at individual level?

1490 

Abstract Submission 

Lars Smolders^1,2, Wouter De Baene³, Geert-Jan Rutten¹, Remco Hofstad², Luc Florack²

¹Elisabeth-Tweesteden Hospital, Tilburg, Netherlands, ²Eindhoven University of Technology, Eindhoven, Netherlands, ³Tilburg University, Tilburg, Netherlands

Lars Smolders, MSc  
Elisabeth-Tweesteden Hospital|Eindhoven University of Technology
Tilburg, Netherlands|Eindhoven, Netherlands

Wouter De Baene, Dr.  
Tilburg University
Tilburg, Netherlands

Geert-Jan Rutten, Dr.  
Elisabeth-Tweesteden Hospital
Tilburg, Netherlands

Remco Hofstad, Prof.  
Eindhoven University of Technology
Eindhoven, Netherlands

Luc Florack, Prof.  
Eindhoven University of Technology
Eindhoven, Netherlands

Understanding how anatomical connections give rise to functional interactions in the brain is a major goal in neuroscience. These ideas have been conceptualized in terms of networks of connections between Regions Of Interest (ROIs), where anatomical white-matter connections are represented by Structural Connectivity (SC) and functional interactions by Functional Connectivity (FC). Recently, several studies have attempted predicting FC from SC at individual level using machine learning methods. If successful, this could have significant benefits for clinical applications. Three such studies (Benkarim et al., 2022; Neudorf et al., 2022; Sarwar et al., 2021) show promising results on large data sets, making them attractive for further development. We examined these studies in detail, to determine whether the models indeed represent meaningful individual-level mappings of SC to FC.

See Figure 1 for a summary. On the data set used in the three studies, comprising MRI images of 1000 Human Connectome Project (HCP) subjects, we calculated SC and empirical FC (eFC) in the same way as Sarwar et al. Briefly, we calculated SC as the streamline count from whole-brain tractography on diffusion-weighted MRI between each ROI pair and calculated FC as the Pearson correlation coefficient between the resting-state fMRI activity of each ROI pair. We used ROIs from the Desikan-Killiany and Schaefer 200 parcel atlas. Using a 10-fold cross-validation workflow, we compared the performance of the presented models to the group average eFC of each training fold (avg-eFC) and to avg-eFC with random noise added to artificially introduce inter-individual differences. We also reproduced analyses demonstrating additional usefulness of predicted FC (pFC) from the authors' models: predicting cognitive performance from pFC (Sarwar et al., 2021), analyzing predictive power of pFC for network centrality measures in eFC (Neudorf et al., 2022) and analyzing differences in performance across Yeo's 7 networks (Benkarim et al., 2022).

Sarwar et al. reported (on Supplementary data) an average individual-level eFC-pFC correlation of 0.7 and showed that pFC was not a group average by reporting an average inter-pFC correlation of 0.7. These values were reproduced by avg-eFC plus noise tuned to 0.1 standard deviation (Figure 2a). A correlation between predicted and actual cognitive score of 0.29 using pFC with SC regressed out was presented. When reproduced by us, we found a correlation of 0.19. Avg-eFC plus noise with SC regressed out achieved a correlation of 0.20. Neudorf et al. showed that pFC explained 56% of variance in eFC at individual level. Avg-eFC explained 55% of variance (Figure 2b). They also showed that centrality measures of pFC explain 81%, 55% and 55% of variance in degree, eigenvector and PageRank centrality respectively. However, Avg-eFC explains 77%, 76% and 77% of variance in these measures. Benkarim et al. presented an average individual-level eFC-pFC correlation of 0.775. Avg-eFC achieved 0.762 (Figure 2c). Finally, differences in performance found by Benkarim et al., i.e. lower performance in default, frontoparietal and limbic networks, were also reproduced by avg-eFC.

We conclude that we cannot ascertain that the models as presented in the three studies represent a meaningful mapping from SC to FC. Firstly, the reported predictive performances do not exceed the performance of avg-eFC. Secondly, analyses into additional usefulness were reproduced by replacing pFC with the avg-eFC plus noise. Although it appears that current methods fail to find a meaningful mapping from SC to FC, we are convinced that this line of work should be continued, taking care to ensure that trained models do not converge to the group average. We believe that taking a step back and re-assessing the construction of SC and FC may yield more informative versions of both, and thus improve the chances of finding a meaningful mapping from structure to function at individual level.

Classification and Predictive Modeling ²

Connectivity (eg. functional, effective, structural) ¹

Diffusion MRI Modeling and Analysis

fMRI Connectivity and Network Modeling

Computational Neuroscience

Computing

Data analysis

FUNCTIONAL MRI

Informatics

Machine Learning

MRI

Tractography

WHITE MATTER IMAGING - DTI, HARDI, DSI, ETC