Higher Criticism—an optimal test for sparse and weak effects in connectomics
Poster No:
1748
Submission Type:
Abstract Submission
Authors:
Andrew Gerlach1, Helmet Karim1, Carmen Andreescu1, Robert Krafty2
Institutions:
1University of Pittsburgh, Pittsburgh, PA, 2Emory University, Atlanta, GA
First Author:
Co-Author(s):
Introduction:
Connectome approaches to analyze MRI data have gained increasing popularity [1]. This comes at the cost of increased dimensionality and sparse/weak effects [2]. Machine learning and graph theory have been popular choices for connectome analysis, though their interpretation is challenging.
Methods:
We introduce Higher Criticism (HC) [3] as a flexible and intuitive tool for connectome analysis. HC is an omnibus test optimal in the rare-and-weak regime that can be applied to mass univariate testing at the edge level. The primary constraints are that the test must be univariate with respect to the connectome and associated with a valid p-value. HC compares the empirical distribution of p-values with the null using a modified Kolmogrov-Smirnov statistic to quantify the deviation (Figure 1). Importantly, this procedure can be carried out at various levels of the connectome (e.g., whole-brain, network, node). We propose that analysis should begin with a whole-brain HC test, which provides no information on localization, but strong motivation for further testing at a more refined level if the result is positive. This procedure can then be repeated for increasing levels of precision, with multiple comparisons correction for only the number of subunits analyzed at a particular level.
We demonstrate HC with data comparing resting state functional connectivity (FC) between 39 never-depressed (ND) and 72 recently remitted depressed (RD) older adults. FC matrices are generated by parcellation of the processed scans with the Schaefer400 atlas [4]. Nodes are grouped into seven canonical Yeo networks [5]. We conduct whole-brain edge-wise independent sample t-tests to differentiate remitters from healthy controls to generate the p-values for HC. We begin with a whole-brain HC test and proceed to network then node levels.
We demonstrate HC with data comparing resting state functional connectivity (FC) between 39 never-depressed (ND) and 72 recently remitted depressed (RD) older adults. FC matrices are generated by parcellation of the processed scans with the Schaefer400 atlas [4]. Nodes are grouped into seven canonical Yeo networks [5]. We conduct whole-brain edge-wise independent sample t-tests to differentiate remitters from healthy controls to generate the p-values for HC. We begin with a whole-brain HC test and proceed to network then node levels.
·Figure 1
Results:
We show how HC can be applied to connectome analysis to detect rare and weak effects, and discuss considerations specific to connectome analysis, specifically: sample size, p-value calculation, and HC variants. Further, we show that HC naturally gives rise to a hierarchical testing procedure through increasing levels of regional specificity that do not require a pre-defined level of analysis and allows for a principled reduction of the multiple comparisons burden.
In our example analysis (Figure 2), we show that the RD and ND groups differ at the whole-brain level. We then use HC to compare each intra- and internetwork block of the FC matrices with Bonferroni correction (m = 28), where we observe 12 network pairs with differing connectivity between RD and ND. Finally, we applied HC to individual rows/columns of the intra- and internetwork blocks to identify nodes with aberrant FC with Bonferroni correction for the number of nodes in each block, resulting in the identification of 154 unique nodes with connectivity differences between groups.
In our example analysis (Figure 2), we show that the RD and ND groups differ at the whole-brain level. We then use HC to compare each intra- and internetwork block of the FC matrices with Bonferroni correction (m = 28), where we observe 12 network pairs with differing connectivity between RD and ND. Finally, we applied HC to individual rows/columns of the intra- and internetwork blocks to identify nodes with aberrant FC with Bonferroni correction for the number of nodes in each block, resulting in the identification of 154 unique nodes with connectivity differences between groups.
·Figure 2
Conclusions:
Higher criticism is a powerful and flexible tool that can be used to exploit the spare and weak signals inherent in connectome-based analysis. HC can naturally be applied in a hierarchical fashion which does not require a strong a priori hypothesis, allows for an optimal balance between regional specificity and power, and reduces the multiple comparisons burden in a principled manner. Thus, interpretable inferences can be made in connectome analysis with reasonable sample sizes. Further, HC allows for relevant, translational interpretations that provides a complementary approach to graph theoretic measures.
We demonstrate application of HC with a dataset comparing remitted depressed and never depressed older adults, showing altered connectivity at the whole-brain, network, and node levels. Given the intuitiveness of HC, we suggest a larger applicability in the neuroimaging field. Further use may lead to methodological improvements accounting for the idiosyncrasies of neuroimaging, especially if the inherent correlation structure can be harnessed.
We demonstrate application of HC with a dataset comparing remitted depressed and never depressed older adults, showing altered connectivity at the whole-brain, network, and node levels. Given the intuitiveness of HC, we suggest a larger applicability in the neuroimaging field. Further use may lead to methodological improvements accounting for the idiosyncrasies of neuroimaging, especially if the inherent correlation structure can be harnessed.
Modeling and Analysis Methods:
Connectivity (eg. functional, effective, structural) 2
Diffusion MRI Modeling and Analysis
fMRI Connectivity and Network Modeling 1
Methods Development
Task-Independent and Resting-State Analysis
Keywords:
Data analysis
Statistical Methods
Univariate
1|2Indicates the priority used for review
Provide references using author date format
2. Marek, S. et al. Reproducible brain-wide association studies require thousands of individuals. Nature 603, 654–660 (2022).
3. Donoho, D. & Jin, J. Higher Criticism for Large-Scale Inference, Especially for Rare and Weak Effects. Statistical Science 30, 1–25 (2015).
4. Schaefer, A. et al. Local-Global Parcellation of the Human Cerebral Cortex from Intrinsic Functional Connectivity MRI. Cereb Cortex 28, 3095–3114 (2018).
5. Thomas Yeo, B. T. et al. The organization of the human cerebral cortex estimated by intrinsic functional connectivity. J Neurophysiol 106, 1125–1165 (2011).