Three anatomical atlases were used to define the node regions of the brain graph following preprocessing of the fMRI images, allowing us to explore reproducibility of the findings across different brain parcellations. These included the Destrieux cortical atlas (148 regions) \cite{Fischl_2004}, the Harvard-Oxford parcellation \cite{Desikan_2006}, including cortical and subcortical structures, and the AAL template (116 regions) \cite{Tzourio_Mazoyer_2002}. For each atlas and each corresponding regions, mean time series were calculated (see Fig. \ref{113850}. Partial correlation between the time series was employed to estimate the strength of the functional connections, yielding a weighted graph representation \cite{Richiardi_2013}. Global efficiency of the networks was explored via the characteristic path length, \(L\), which corresponds to the average path length \(l(s)\) across all nodes \(s\) \cite{Rubinov_2010}. This measure was calculated separately for positive, negative and absolute weights of the estimated partial correlation networks, as there is no consensus with regards to which of these is most discriminative.

Apart from looking at the association between the number of rich club nodes affected by stroke (\(N_{RC}\)), and the outcome measures collected at the early and late follow-ups, we devised four different predictive linear regression models. The baseline model,