Figure 2
Preserved features in latent space: (A) Calculation of edge time series by means of Hadamard product at each time point. The highest top amplitude frames (top %10 of co-fluctuation root-sum-squared) were selected. (B) The collective co-fluctuations of brain regions were estimated as the root sum square (RSS) of co-fluctuations between all pairs of brain regions (edges) at every given time point. The peaks that occur at the same time point in the source and latent space were labeled as “peak hits” and used to calculate the “peak agreement”. This amount was used as an indicator of preserved dynamics. (C) The amount of peak agreement was calculated for each dimension revealing dimension 2 as the lowest. There was a significant difference between dimension 5 and 6 (p < 0.01). Following dimension 6, the level of agreement stabilizes. (D) The original FC was contrasted with the FC created only by the timepoints containing peaks, showing a correlation of R = 0.93. The FC strength (average of each node) was projected onto surface in order to observe the spatial pattern similarity. (E) The correlation between the FC of the timepoints containing peaks and the complete FC was performed for all the subjects in order to characterize the variability across subjects. (F) We compared the distribution of the FCD of healthy controls and stroke patients. The difference observed in source space (p < 0.01) increased when observing the latent representation (p < 0.01). (G) Edge metastability was compared between controls and stroke patients showing no significant difference in the source space (p = 0.9) in contrast to the latent space (p < 0.01). (H) Level of modularity was significantly different in controls and stroke patients both in source (p < 0.01) and latent space (p < 0.01). (I) Level of functional complexity was significantly different in controls and stroke patients both in source (p < 0.01) and latent space (p < 0.01).
