Data-driven Topological Filtering based on Orthogonal Minimal Spanning Trees: Application to Multi-Group MEG Resting-State Connectivity

Abstract

In the present study a novel data-driven topological filtering technique is introduced to derive the backbone of functional brain networks relying on orthogonal minimal spanning trees (OMST).The method aims to identify the essential functional connections to ensure optimal information flow via the objective criterion of global efficiency minus the cost of surviving connections. TheOMST technique was applied to multichannel, resting-state neuromagnetic recordings from four groups of participants: healthy adults (n=50), adults who have suffered mild traumatic brain injury (n=30), typically developing children (n=27), and reading-disabled children (n=25).Weighted interactions between network nodes (sensors) were computed using an integrated approach of dominant intrinsic coupling modes based on two alternative metrics (symbolic mutual information and phase lag index), resulting in excellent discrimination of individual cases according to their group membership. Classification results using OMST-derived functional networks were clearly superior to results using either relative power spectrum features or functional networks derived through the conventional minimal spanning trees algorithm.

Fronto‐Parietal Subnetworks Flexibility Compensates For Cognitive Decline Due To Mental Fatigue

“…In this study, we focused on the scale of 2, which corresponds to the frequency range of 0.06 ‐ 0.125 Hz. The absolute values of Fisher's Z‐transformed correlation coefficients were then thresholded with the data‐driven topological filtering scheme based on optimization of the global cost‐efficiency of the network (Bassett et al, ) and the OMSTs (Dimitriadis, Antonakakis, et al, ; Dimitriadis, Salis, et al, ) (see Appendix A for more details). The obtained topological undirected weighted graphs were used for the subsequent community structure analysis.…”

“…We targeted the short‐term effects of prolonged task performance, which were considered to be due to mental fatigue and persisted for at least several minutes (Breckel et al, ), rather than long‐term training effects. Individual functional networks at pre‐ and posttask runs were constructed with cutting‐edge techniques including maximum overlap discrete wavelet transform (MODWT), Ledoit‐Wolf shrinkage covariance estimator (Ledoit & Wolf, ) and a data‐driven topological filtering scheme based on the orthogonal minimum spanning trees (OMSTs) (Dimitriadis, Antonakakis, Simos, Fletcher, & Papanicolaou, ; Dimitriadis, Salis, Tarnanas, & Linden, ), which yielded reproducible networks. The subnetwork metrics and conventional metrics for the whole networks were introduced as explanatory variables to regression analyses to examine which metrics were effective in discriminating pre‐ and posttask periods and in predicting extents of individual cognitive decline.…”

Alterations in Dynamic Spontaneous Network Microstates in Mild Traumatic Brain Injury: A MEG Beamformed Dynamic Connectivity Analysis

“…In addition to statistical filtering, a data-based topological connection-cutting was applied based on a recently suggested procedure (Dimitriadis et al, 2017a, b). The so-called Orthogonal Minimal Spanning Trees (OMST 1 ) was performed to uncover the entire structure of the most dominant paths within every quasi-static FCG.…”

“…The so-called Orthogonal Minimal Spanning Trees (OMST 1 ) was performed to uncover the entire structure of the most dominant paths within every quasi-static FCG. The the OMST procedure was first enhanced from the notion that the full connected FCG can be extracted to an acyclic FCG or MST of minimum cost from the root node to leaf node without changing the strong connections strong from the weak connections (Dimitriadis et al, 2017a, b). Subsequently, the global efficiency of the specific MST was optimized preserving same total cost among within the connections (Dimitriadis et al, 2017b, d).…”

“…The within brain interactions were modeled using beamformed MEG source activation and DFC among ninety atlas-based brain areas using a template MRI. Each quasi-static FC was determined by the imaginary content of the phase locking value (iPLV) and it was filtered statistically and topologically (Dimitriadis et al, 2017a,b) for the reduction of spurious connections. Subsequently, we coded the estimated time-varying network activity into prototypical network microstates or FCμstates (Dimitriadis et al, 2013a, 2018a).…”