The dynamic functional connectivity fingerprint of high-grade gliomas

Resting state fMRI has been used in many studies to investigate the impact of brain tumours on functional connectivity (FC). However, these studies have so far assumed that FC is stationary, disregarding the fact that the brain fluctuates over dynamic states. Here we utilised resting state fMRI data from 33 patients with high-grade gliomas and 33 healthy controls to examine the dynamic interplay between resting-state networks and to gain insights into the impact of brain tumours on functional dynamics. By employing Hidden Markov Models, we demonstrated that functional dynamics persist even in the presence of a high-grade glioma, and that patients exhibited a global decrease of connections strength, as well as of network segregation. Furthermore, through a multivariate analysis, we demonstrated that patients’ cognitive scores are highly predictive of pathological dynamics, thus supporting our hypothesis that functional dynamics could serve as valuable biomarkers for better understanding the traits of high-grade gliomas.


Participants
The demographics and clinical data of the patients are reported in Table 1. Figure 1 shows the frequency maps of the tumour in the patient population.

Hidden Markov Model -setup
The model was fitted for different model orders and, in particular, from 2 to 15 states. Table 2 reports the values of the indices evaluated for the choice of the optimal order (K). While the free-energy (FE) showed a decreasing trend as the number of states increased, the average log-likelihood (avLL) reached the maximum for K=4, followed by K=9,6. Considering only these three orders, the coefficients of variation (CVs) suggested K=6 followed by K=9 as the best. Furthermore, the CVs obtained within each state, were, on average, lower for K=6 compared to K=9. Therefore, the number of states was set to K=6.  Table 2: Three metrics, employed for the choice of the best HMM order, are reported: the free-energy (FE), the average log-likelihood (avLL), and the coefficients of variation (CVs). Figure 2 reports the transition probabilities among states, separately for patients and healthy controls.

Supplementary Figure 2:
Transition probabilities from one state (y-axis) to another (x-axis) for the two groups of healthy controls and patients separately. The table reports the Jaccard similarity values (Jaccard) between each state modular matrix. This index ranges between 0 and 1, where a similarity of 1 means that the states share the same modular organization.

Methods
To compare the results obtained through HMM with the gold-standard approach employed in the literature for dynamic connectivity analyses, we performed a sliding window followed by clustering analysis (SW) 1 . In details, we chose a window size of 73 TR (92 seconds) and a step size of 1 TR. Thus, for each subject we obtained 584 windowed FC matrices. For each windowed FC, only the values in the upper triangular part of the matrix were retained after z-Fisher transformation. Considering both the two groups of subjects, we thus obtained a matrix of dimensions 66 subjects × 584 windows × 990 correlation values. All subjects were then concatenated prior to performing the clustering analysis, thus obtaining a matrix of dimensions 38544×990. Then, a K-means clustering was performed on the windowed FC matrices (Euclidean distance, 50 repetitions, number of clusters from 2 to 10). The Silhouette criterion 2 was employed for the choice of the optimal cluster size. For each cluster centroid the corresponding FC matrix was created and on the same analyses performed on the brain states FC matrices were used. In particular, the FO in each centroid was computed for each subject (FO_SW) and then to compare the FO_SW among the two groups, a Wilcoxon's rank sum test followed by multiple comparison correction (α=0.05) was applied. The graph-based analyses and the same statistical tests described before were repeated on the centroids FC matrices. To assess whether the two approaches (i.e., HMM and SW) could give comparable results in terms of connectivity and graph metrics associated to the states or clusters centroids, we matched each HMM state with a cluster centroid on the basis of the associated FC matrix. Specifically, the Pearson's correlation and the structural similarity index between the states FC matrices and the clusters centroids FC matrices were computed.

Results
The optimal cluster size resulted to be equal to six, employing the Silhouette criterion, thus six centroids FC matrices were characterized (Supplementary Figure 3). When we compared the FO in different clusters between the two groups, we found statistically significant differences only in two clusters over six. Cluster 5 was mostly populated by patients, while cluster 6 by HCs (Supplementary Figure 4). The match between FC matrices obtained with the two approaches (HMM and SW), evaluated through the Pearson's correlation and the structural similarity index (Supplementary Table 4 and 5), pointed out that the pathological state S5 could be clearly associated to cluster 5, while the healthy state S2 and S3 to cluster 2 and 3. For the remaining states, only a weak association was registered. Interestingly, in cluster 5, similar to the pathological S5, a statistically significant decrease in local efficiency and clustering coefficient, but not in the degree and betweenness centrality, was found (Supplementary Figure 5). Thus, while through the HMM analysis we were able to find two pathological brain states, through the SW analysis we could only find one pathological cluster, which probably encapsulated the connectivity features of both S5 and S6. Thus, in contrast to HMM-based analysis, this approach was less sensitive in separating patients from controls.