Evaluating Regularized Modeling Methods for Calcuating Functional Connectivity
McKenzie Hagen, University of Washington
Introduction: Functional connectomes describe the statistical relation between spatial distinct brain signals, and are used to investigate brain dynamics and how they relate to behavior and cognition. Functional MRI (fMRI) researchers typically construct connectomes by calculating the Pearson correlations between every pair of brain signals. Recent research demonstrates the drawbacks of this measurement, such as high correlation with motion.(1) Here, we present and evaluate alternative modeling methods for constructing connectomes from fMRI data.
Methods: Using a large, high quality fMRI dataset we compared connectomes constructed using Pearson correlations, LASSO regularized models, and the Union of Intersections (UofI) algorithm(2). Comparison criteria include characteristics that are theoretically, psychometrically, and empirically desired for connectome analyses, such as graph sparsity, test-retest reliability and participant identification (“fingerprinting”).
Results: Preliminary analyses show that connectomes calculated by UoI and LASSO outperform or match Pearson connectomes on selected criteria. UofI and LASSO connectomes are necessarily sparser than Pearson correlation connectomes. All connectomes compared are moderately reliable between timepoints, and highly accurate for participant fingerprinting.
Conclusion: Pearson correlation connectomes are pervasive in fMRI research, despite their problematic characteristics. Alternative connectome calculation methods like LASSO and UofI match or outperform Pearson connectomes on several key criteria, demonstrating their suitability for studying brain dynamics. Future research should investigate how well these methods perform for brain-behavior analyses as compared to Pearson connectomes.
Mahadevan, A. S. et al., NeuroImage (2021)
Sachdeva, P. S., Bhattacharyya, S. & Bouchard, K. E. 2019 (EMBC) (2019).