A novel approach to understanding Parkinsonian cognitive decline using minimum spanning trees, edge cutting, and magnetoencephalography

Olivier B Simon,Donald C Rojas,Debashis Ghosh,Isabelle Buard,Benzi M Kluger,Samantha K Holden

doi:10.1038/s41598-021-99167-2

Abstract

Graph theory-based approaches are efficient tools for detecting clustering and group-wise differences in high-dimensional data across a wide range of fields, such as gene expression analysis and neural connectivity. Here, we examine data from a cross-sectional, resting-state magnetoencephalography study of 89 Parkinson’s disease patients, and use minimum-spanning tree (MST) methods to relate severity of Parkinsonian cognitive impairment to neural connectivity changes. In particular, we implement the two-sample multivariate-runs test of Friedman and Rafsky (Ann Stat 7(4):697–717, 1979) and find it to be a powerful paradigm for distinguishing highly significant deviations from the null distribution in high-dimensional data. We also generalize this test for use with greater than two classes, and show its ability to localize significance to particular sub-classes. We observe multiple indications of altered connectivity in Parkinsonian dementia that may be of future use in diagnosis and prediction.

Highlights

We explore the possibilities of minimum spanning tree (MST)-based methods, the multivariate runs test, to detect group differences in medical information by performing connectivity analyses on cross-sectional resting-state magnetoencephalography (MEG) data drawn from three classes of Parkinson’s disease (PD) patients—normal cognition (PD-NC), mild cognitive impairment (PD-Mild Cognitive Impairment (MCI)) and full dementia (PDD)
Aggregated over all conditions, we found nodes from the PD-NC and MCI classes were most likely to share edges in the 3-class MST, while PD-NC and Parkinson’s disease dementia (PDD) were much less likely; the number of shared edges between PD-MCI and PDD were intermediate (Fig. 5)
We found strong indications in all bands that connectivity becomes considerably less variable between patients with dementia progression, while for lower-frequency bands there is a partial recovery of this variability going from PD-MCI to PDD

Summary

Introduction

More broadly, network science, provides a natural mathematical framework for the representation and analysis of complex networks[1,3,4], giving it clear applicability for studies of brain connectivity. In such studies, brain regions of interest are represented as "nodes" of a graph structure, while significant information flow between the nodes is represented as connecting "edges". One approach to the study of graph connectivity is to threshold the connectivity values so that connections below a certain strength are excluded from the graph This introduces a subjective parameter into the analysis, which may dramatically affect r esults[7].

Methods

Results

Conclusion